network calculus - Computer Engineering and Networks Laboratory

NETWORK CALCULUS A Theory of Deterministic Queuing Systems for the Internet J EAN -Y VES L E B OUDEC PATRICK T HIRAN

Online Version of the Book Springer Verlag - LNCS 2050 Version May 10, 2004

2

A Annelies A Joana, Maëlle, Audraine et Elias A ma mère —- JL A mes parents —- PT

Pour e´ viter les grumeaux Qui encombrent les réseaux Il fallait, c’est compliqué, Maˆıtriser les seaux percés Branle-bas dans les campus On pourra dorénavant Calculer plus simplement Grâce a` l’algèbre Min-Plus Foin des obscures astuces Pour estimer les délais Et la gigue des paquets Place a` “Network Calculus” —- JL

vi Summary of Changes 2002 Jan 14, JL Chapter 2: added a better coverage of GR nodes, in particular equivalence with service curve. Fixed bug in Proposition 1.4.1 2002 Jan 16, JL Chapter 6: M. Andrews brought convincing proof that conjecture 6.3.1 is wrong. Redesigned Chapter 6 to account for this. Removed redundancy between Section 2.4 and Chapter 6. Added SETF to Section 2.4 2002 Feb 28, JL Bug fixes in Chapter 9 2002 July 5, JL Bug fixes in Chapter 6; changed format for a better printout on most usual printers. 2003 June 13, JL Added concatenation properties of non-FIFO GR nodes to Chapter 2. Major upgrade of Chapter 7. Reorganized Chapter 7. Added new developments in Diff Serv. Added properties of PSRG for non-FIFO nodes. 2003 June 25, PT Bug fixes in chapters 4 and 5. 2003 Sept 16, JL Fixed bug in proof of theorem 1.7.1, proposition 3. The bug was discovered and brought to our attention by François Larochelle. 1 1 2004 Jan 7, JL Bug fix in Proposition 2.4.1 (ν > h−1 instead of ν < h−1 ) 2004, May 10, JL Typo fixed in Definition 1.2.4 (thanks to Richard Bradford)

Contents Introduction

xiii

I A First Course in Network Calculus

1

1 Network Calculus

3

1.1

1.2

1.3

1.4

1.5

1.6

Models for Data Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

1.1.1

Cumulative Functions, Discrete Time versus Continuous Time Models . . . . . . . .

3

1.1.2

Backlog and Virtual Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

1.1.3

Example: The Playout Buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6

Arrival Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.2.1

Definition of an Arrival Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.2.2

Leaky Bucket and Generic Cell Rate Algorithm . . . . . . . . . . . . . . . . . . . . 10

1.2.3

Sub-additivity and Arrival Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.2.4

Minimum Arrival Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Service Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.3.1

Definition of Service Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.3.2

Classical Service Curve Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

Network Calculus Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 1.4.1

Three Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1.4.2

Are the Bounds Tight ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.4.3

Concatenation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

1.4.4

Improvement of Backlog Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

Greedy Shapers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 1.5.1

Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.5.2

Input-Output Characterization of Greedy Shapers . . . . . . . . . . . . . . . . . . . 31

1.5.3

Properties of Greedy Shapers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

Maximum Service Curve, Variable and Fixed Delay . . . . . . . . . . . . . . . . . . . . . . 34 1.6.1

Maximum Service Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

1.6.2

Delay from Backlog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.6.3

Variable versus Fixed Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 vii

CONTENTS

viii 1.7

1.8

1.9

Handling Variable Length Packets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 1.7.1

An Example of Irregularity Introduced by Variable Length Packets . . . . . . . . . . 40

1.7.2

The Packetizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

1.7.3

A Relation between Greedy Shaper and Packetizer . . . . . . . . . . . . . . . . . . 45

1.7.4

Packetized Greedy Shaper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

Effective Bandwidth and Equivalent Capacity . . . . . . . . . . . . . . . . . . . . . . . . . 53 1.8.1

Effective Bandwidth of a Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

1.8.2

Equivalent Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

1.8.3

Example: Acceptance Region for a FIFO Multiplexer . . . . . . . . . . . . . . . . . 55

Proof of Theorem 1.4.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

1.10 Bibliographic Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 1.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2 Application to the Internet 2.1

2.2

2.3

2.4

67

GPS and Guaranteed Rate Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.1.1

Packet Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

2.1.2

GPS and a Practical Implementation (PGPS) . . . . . . . . . . . . . . . . . . . . . 68

2.1.3

Guaranteed Rate (GR) Nodes and the Max-Plus Approach . . . . . . . . . . . . . . 70

2.1.4

Concatenation of GR nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

2.1.5

Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

The Integrated Services Model of the IETF . . . . . . . . . . . . . . . . . . . . . . . . . . 75 2.2.1

The Guaranteed Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

2.2.2

The Integrated Services Model for Internet Routers . . . . . . . . . . . . . . . . . . 75

2.2.3

Reservation Setup with RSVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

2.2.4

A Flow Setup Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2.2.5

Multicast Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

2.2.6

Flow Setup with ATM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

Schedulability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 2.3.1

EDF Schedulers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

2.3.2

SCED Schedulers [73] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

2.3.3

Buffer Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Application to Differentiated Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 2.4.1

Differentiated Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

2.4.2

An Explicit Delay Bound for EF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

2.4.3

Bounds for Aggregate Scheduling with Dampers . . . . . . . . . . . . . . . . . . . 93

2.4.4

Static Earliest Time First (SETF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

2.5

Bibliographic Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

2.6

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

CONTENTS

ix

II Mathematical Background

101

3 Basic Min-plus and Max-plus Calculus

103

3.1

Min-plus Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 3.1.1

Infimum and Minimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

3.1.2

Dioid (R ∪ {+∞}, ∧, +) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

3.1.3

A Catalog of Wide-sense Increasing Functions . . . . . . . . . . . . . . . . . . . . 105

3.1.4

Pseudo-inverse of Wide-sense Increasing Functions . . . . . . . . . . . . . . . . . . 108

3.1.5

Concave, Convex and Star-shaped Functions . . . . . . . . . . . . . . . . . . . . . 109

3.1.6

Min-plus Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

3.1.7

Sub-additive Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

3.1.8

Sub-additive Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

3.1.9

Min-plus Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

3.1.10 Representation of Min-plus Deconvolution by Time Inversion . . . . . . . . . . . . 125 3.1.11 Vertical and Horizontal Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 3.2

3.3

Max-plus Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 3.2.1

Max-plus Convolution and Deconvolution . . . . . . . . . . . . . . . . . . . . . . . 129

3.2.2

Linearity of Min-plus Deconvolution in Max-plus Algebra . . . . . . . . . . . . . . 129

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

4 Min-plus and Max-Plus System Theory 4.1

131

Min-Plus and Max-Plus Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.1.1

Vector Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.1.2

Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

4.1.3

A Catalog of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

4.1.4

Upper and Lower Semi-Continuous Operators . . . . . . . . . . . . . . . . . . . . . 134

4.1.5

Isotone Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

4.1.6

Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

4.1.7

Causal Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

4.1.8

Shift-Invariant Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

4.1.9

Idempotent Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

4.2

Closure of an Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

4.3

Fixed Point Equation (Space Method) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 4.3.1

Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

4.3.2

Examples of Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

4.4

Fixed Point Equation (Time Method) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

4.5

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

CONTENTS

x

III A Second Course in Network Calculus

153

5 Optimal Multimedia Smoothing

155

5.1

Problem Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

5.2

Constraints Imposed by Lossless Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . 156

5.3

Minimal Requirements on Delays and Playback Buffer . . . . . . . . . . . . . . . . . . . . 157

5.4

Optimal Smoothing Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 5.4.1

Maximal Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

5.4.2

Minimal Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

5.4.3

Set of Optimal Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

5.5

Optimal Constant Rate Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

5.6

Optimal Smoothing versus Greedy Shaping . . . . . . . . . . . . . . . . . . . . . . . . . . 163

5.7

Comparison with Delay Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

5.8

Lossless Smoothing over Two Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

5.9

5.8.1

Minimal Requirements on the Delays and Buffer Sizes for Two Networks . . . . . . 169

5.8.2

Optimal Constant Rate Smoothing over Two Networks . . . . . . . . . . . . . . . . 171


6 Aggregate Scheduling

175

6.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

6.2

Transformation of Arrival Curve through Aggregate Scheduling . . . . . . . . . . . . . . . 176

6.3

6.4

6.2.1

Aggregate Multiplexing in a Strict Service Curve Element . . . . . . . . . . . . . . 176

6.2.2

Aggregate Multiplexing in a FIFO Service Curve Element . . . . . . . . . . . . . . 177

6.2.3

Aggregate Multiplexing in a GR Node . . . . . . . . . . . . . . . . . . . . . . . . . 181

Stability and Bounds for a Network with Aggregate Scheduling . . . . . . . . . . . . . . . . 181 6.3.1

The Issue of Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

6.3.2

The Time Stopping Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

Stability Results and Explicit Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 6.4.1

The Ring is Stable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

6.4.2

Explicit Bounds for a Homogeneous ATM Network with Strong Source Rate Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

6.5


6.6

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

7 Adaptive and Packet Scale Rate Guarantees

197

7.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

7.2

Limitations of the Service Curve and GR Node Abstractions . . . . . . . . . . . . . . . . . 197

7.3

Packet Scale Rate Guarantee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 7.3.1

Definition of Packet Scale Rate Guarantee . . . . . . . . . . . . . . . . . . . . . . . 198

7.3.2

Practical Realization of Packet Scale Rate Guarantee . . . . . . . . . . . . . . . . . 202

CONTENTS 7.3.3 7.4

7.5

xi Delay From Backlog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

Adaptive Guarantee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 7.4.1

Definition of Adaptive Guarantee . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

7.4.2

Properties of Adaptive Guarantees . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

7.4.3

PSRG and Adaptive Service Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Concatenation of PSRG Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 7.5.1

Concatenation of FIFO PSRG Nodes . . . . . . . . . . . . . . . . . . . . . . . . . 206

7.5.2

Concatenation of non FIFO PSRG Nodes . . . . . . . . . . . . . . . . . . . . . . . 207

7.6

Comparison of GR and PSRG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

7.7

Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 7.7.1

Proof of Lemma 7.3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

7.7.2

Proof of Theorem 7.3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

7.7.3

Proof of Theorem 7.3.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

7.7.4

Proof of Theorem 7.3.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

7.7.5

Proof of Theorem 7.4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

7.7.6

Proof of Theorem 7.4.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

7.7.7

Proof of Theorem 7.4.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

7.7.8

Proof of Theorem 7.4.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

7.7.9

Proof of Theorem 7.5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

7.7.10 Proof of Proposition 7.5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 7.8


7.9

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

8 Time Varying Shapers

225

8.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

8.2

Time Varying Shapers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

8.3

Time Invariant Shaper with Initial Conditions . . . . . . . . . . . . . . . . . . . . . . . . . 227 8.3.1

Shaper with Non-empty Initial Buffer . . . . . . . . . . . . . . . . . . . . . . . . . 227

8.3.2

Leaky Bucket Shapers with Non-zero Initial Bucket Level . . . . . . . . . . . . . . 228

8.4

Time Varying Leaky-Bucket Shaper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

8.5


9 Systems with Losses 9.1

231

A Representation Formula for Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 9.1.1

Losses in a Finite Storage Element . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

9.1.2

Losses in a Bounded Delay Element . . . . . . . . . . . . . . . . . . . . . . . . . . 233

9.2

Application 1: Bound on Loss Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

9.3

Application 2: Bound on Losses in Complex Systems . . . . . . . . . . . . . . . . . . . . . 235 9.3.1

Bound on Losses by Segregation between Buffer and Policer . . . . . . . . . . . . . 235

CONTENTS

xii 9.3.2

Bound on Losses in a VBR Shaper . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

9.4

Skohorkhod’s Reflection Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

9.5


I NTRODUCTION

W HAT THIS B OOK IS A BOUT Network Calculus is a set of recent developments that provide deep insights into flow problems encountered in networking. The foundation of network calculus lies in the mathematical theory of dioids, and in particular, the Min-Plus dioid (also called Min-Plus algebra). With network calculus, we are able to understand some fundamental properties of integrated services networks, window flow control, scheduling and buffer or delay dimensioning. This book is organized in three parts. Part I (Chapters 1 and 2) is a self contained, first course on network calculus. It can be used at the undergraduate level or as an entry course at the graduate level. The prerequisite is a first undergraduate course on linear algebra and one on calculus. Chapter 1 provides the main set of results for a first course: arrival curves, service curves and the powerful concatenation results are introduced, explained and illustrated. Practical definitions such as leaky bucket and generic cell rate algorithms are cast in their appropriate framework, and their fundamental properties are derived. The physical properties of shapers are derived. Chapter 2 shows how the fundamental results of Chapter 1 are applied to the Internet. We explain, for example, why the Internet integrated services internet can abstract any router by a ratelatency service curve. We also give a theoretical foundation to some bounds used for differentiated services. Part II contains reference material that is used in various parts of the book. Chapter 3 contains all first level mathematical background. Concepts such as min-plus convolution and sub-additive closure are exposed in a simple way. Part I makes a number of references to Chapter 3, but is still self-contained. The role of Chapter 3 is to serve as a convenient reference for future use. Chapter 4 gives advanced min-plus algebraic results, which concern fixed point equations that are not used in Part I. Part III contains advanced material; it is appropriate for a graduate course. Chapter 5 shows the application of network calculus to the determination of optimal playback delays in guaranteed service networks; it explains how fundamental bounds for multimedia streaming can be determined. Chapter 6 considers systems with aggregate scheduling. While the bulk of network calculus in this book applies to systems where schedulers are used to separate flows, there are still some interesting results that can be derived for such systems. Chapter 7 goes beyond the service curve definition of Chapter 1 and analyzes adaptive guarantees, as they are used by the Internet differentiated services. Chapter 8 analyzes time varying shapers; it is an extension of the fundamental results in Chapter 1 that considers the effect of changes in system parameters due to adaptive methods. An application is to renegotiable reserved services. Lastly, Chapter 9 tackles systems with losses. The fundamental result is a novel representation of losses in flow systems. This can be used to bound loss or congestion probabilities in complex systems. Network calculus belongs to what is sometimes called “exotic algebras” or “topical algebras”. This is a set of mathematical results, often with high description complexity, that give insights into man-made systems xiii

INTRODUCTION

xiv

such as concurrent programs, digital circuits and, of course, communication networks. Petri nets fall into this family as well. For a general discussion of this promising area, see the overview paper [35] and the book [28]. We hope to convince many readers that there is a whole set of largely unexplored, fundamental relations that can be obtained with the methods used in this book. Results such as “shapers keep arrival constraints” or “pay bursts only once”, derived in Chapter 1 have physical interpretations and are of practical importance to network engineers. All results here are deterministic. Beyond this book, an advanced book on network calculus would explore the many relations between stochastic systems and the deterministic relations derived in this book. The interested reader will certainly enjoy the pioneering work in [28] and [11]. The appendix contains an index of the terms defined in this book.

N ETWORK C ALCULUS , A S YSTEM T HEORY FOR C OMPUTER N ETWORKS In the rest of this introduction we highlight the analogy between network calculus and what is called “system theory”. You may safely skip it if you are not familiar with system theory. Network calculus is a theory of deterministic queuing systems found in computer networks. It can also be viewed as the system theory that applies to computer networks. The main difference with traditional system theory, as the one that was so successfully applied to design electronic circuits, is that here we consider another algebra, where the operations are changed as follows: addition becomes computation of the minimum, multiplication becomes addition. Before entering the subject of the book itself, let us briefly illustrate some of the analogies and differences between min-plus system theory, as applied in this book to communication networks, and traditional system theory, applied to electronic circuits. Let us begin with a very simple circuit, such as the RC cell represented in Figure 1. If the input signal is the voltage x(t) ∈ R, then the output y(t) ∈ R of this simple circuit is the convolution of x by the impulse response of this circuit, which is here h(t) = exp(−t/RC)/RC for t ≥ 0:

t

y(t) = (h ⊗ x)(t) =

h(t − s)x(s)ds.

0

Consider now a node of a communication network, which is idealized as a (greedy) shaper. A (greedy) shaper is a device that forces an input flow x(t) to have an output y(t) that conforms to a given set of rates according to a traffic envelope σ (the shaping curve), at the expense of possibly delaying bits in the buffer. Here the input and output ‘signals’ are cumulative flow, defined as the number of bits seen on the data flow in time interval [0, t]. These functions are non-decreasing with time t. Parameter t can be continuous or discrete. We will see in this book that x and y are linked by the relation y(t) = (σ ⊗ x)(t) =

inf

s∈R such that 0≤s≤t

{σ(t − s) + x(s)} .

This relation defines the min-plus convolution between σ and x. Convolution in traditional system theory is both commutative and associative, and this property allows to easily extend the analysis from small to large scale circuits. For example, the impulse response of the circuit of Figure 2(a) is the convolution of the impulse responses of each of the elementary cells: h(t) = (h1 ⊗ h2 )(t) =

0

t

h1 (t − s)h2 (s)ds.

xv

5

[W

\W

&

(a) [W

\W

σ (b)

Figure 1: An RC circuit (a) and a greedy shaper (b), which are two elementary linear systems in their respective algebraic structures. The same property applies to greedy shapers, as we will see in Chapter 1. The output of the second shaper of Figure 2(b) is indeed equal to y(t) = (σ ⊗ x)(t), where σ(t) = (σ1 ⊗ σ2 )(t) =

inf


{σ1 (t − s) + σ2 (s)} .

This will lead us to understand the phenomenon known as “pay burst only once” already mentioned earlier in this introduction.

[W

K

\W

K

(a) [W

σ1

σ2

\W

(b)

Figure 2: The impulse response of the concatenation of two linear circuit is the convolution of the individual impulse responses (a), the shaping curve of the concatenation of two shapers is the convolution of the individual shaping curves (b).

There are thus clear analogies between “conventional” circuit and system theory, and network calculus. There are however important differences too. A first one is the response of a linear system to the sum of the inputs. This is a very common situation, in both electronic circuits (take the example of a linear low-pass filter used to clean a signal x(t) from additive

INTRODUCTION

xvi

noise n(t), as shown in Figure 3(a)), and in computer networks (take the example a link of a buffered node with output link capacity C, where one flow of interest x(t) is multiplexed with other background traffic n(t), as shown in Figure 3(b)).

[W

K

\WRWW

\QW

QW

\W

(a) QW &

\WRWW

[W (b)

Figure 3: The response ytot (t) of a linear circuit to the sum of two inputs x + n is the sum of the individual responses (a), but the response ytot (t) of a greedy shaper to the aggregate of two input flows x + n is not the sum of the individual responses (b).

Since the electronic circuit of Figure 3(a) is a linear system, the response to the sum of two inputs is the sum of the individual responses to each signal. Call y(t) the response of the system to the pure signal x(t), yn (t) the response to the noise n(t), and ytot (t) the response to the input signal corrupted by noise x(t) + n(t). Then ytot (t) = y(t) + yn (t). This useful property is indeed exploited to design the optimal linear system that will filter out noise as much as possible. If traffic is served on the outgoing link as soon as possible in the FIFO order, the node of Figure 3(b) is equivalent to a greedy shaper, with shaping curve σ(t) = Ct for t ≥ 0. It is therefore also a linear system, but this time in min-plus algebra. This means that the response to the minimum of two inputs is the minimum of the responses of the system to each input taken separately. However, this also mean that the response to the sum of two inputs is no longer the sum of the responses of the system to each input taken separately, because now x(t) + n(t) is a nonlinear operation between the two inputs x(t) and n(t): it plays the role of a multiplication in conventional system theory. Therefore the linearity property does unfortunately not apply to the aggregate x(t) + n(t). As a result, little is known on the aggregate of multiplexed flows. Chapter 6 will learn us some new results and problems that appear simple but are still open today. In both electronics and computer networks, nonlinear systems are also frequently encountered. They are however handled quite differently in circuit theory and in network calculus. Consider an elementary nonlinear circuit, such as the BJT amplifier circuit with only one transistor, shown in Figure 4(a). Electronics engineers will analyze this nonlinear circuit by first computing a static operating point y for the circuit, when the input x is a fixed constant voltage (this is the DC analysis). Next they will linearize the nonlinear element (i.e the transistor) around the operating point, to obtain a so-called small signal model, which a linear model of impulse response h(t) (this is the AC analysis). Now xlin (t) = x(t) − x is a time varying function of time within a small range around x , so that ylin (t) = y(t) − y is indeed approximately given by ylin (t) ≈ (h ⊗ xlin )(t). Such a model is shown on Figure 4(b). The difficulty of a thorough nonlinear analysis is thus bypassed by restricting the input signal in a small range around the operating point. This allows to use a linearized model whose accuracy is sufficient to evaluate performance measures of interest, such as the gain of the amplifier.

xvii

9FF

\W

[W

[OLQW

\OLQW

(a)

Buffered window flow Controller

(b)

Network β Π

[W

(c)

\W

β

[W

\OLQW

(d)

Figure 4: An elementary nonlinear circuit (a) replaced by a (simplified) linear model for small signals (b), and a nonlinear network with window flow control (c) replaced by a (worst-case) linear system (d).

In network calculus, we do not decompose inputs in a small range time-varying part and another large constant part. We do however replace nonlinear elements by linear systems, but the latter ones are now a lower bound of the nonlinear system. We will see such an example with the notion of service curve, in Chapter 1: a nonlinear system y(t) = Π(x)(t) is replaced by a linear system ylin (t) = (β ⊗ x)(t), where β denotes this service curve. This model is such that ylin (t) ≤ y(t) for all t ≥ 0, and all possible inputs x(t). This will also allow us to compute performance measures, such as delays and backlogs in nonlinear systems. An example is the window flow controller illustrated in Figure 4(c), which we will analyze in Chapter 4. A flow x is fed via a window flow controller in a network that realizes some mapping y = Π(x). The window flow controller limits the amount of data admitted in the network in such a way that the total amount of data in transit in the network is always less than some positive number (the window size). We do not know the exact mapping Π, we assume that we know one service curve β for this flow, so that we can replace the nonlinear system of Figure 4(c) by the linear system of Figure 4(d), to obtain deterministic bounds on the end-to-end delay or the amount of data in transit. The reader familiar with traditional circuit and system theory will discover many other analogies and differences between the two system theories, while reading this book. We should insist however that no prerequisite in system theory is needed to discover network calculus as it is exposed in this book.

ACKNOWLEDGEMENT We gratefully acknowledge the pioneering work of Cheng-Shang Chang and René Cruz; our discussions with them have influenced this text. We thank Anna Charny, Silvia Giordano, Olivier Verscheure, Frédéric

xviii

INTRODUCTION

Worm, Jon Bennett, Kent Benson, Vicente Cholvi, William Courtney, Juan Echaguë, Felix Farkas, Gérard Hébuterne, Milan Vojnović and Zhi-Li Zhang for the fruitful collaboration. The interaction with Rajeev Agrawal, Matthew Andrews, François Baccelli, Guillaume Urvoy and Lothar Thiele is acknowledged with thanks. We are grateful to Holly Cogliati for helping with the preparation of the manuscript.

PART I

A F IRST C OURSE IN N ETWORK C ALCULUS

1

C HAPTER 1

N ETWORK C ALCULUS In this chapter we introduce the basic network calculus concepts of arrival, service curves and shapers. The application given in this chapter concerns primarily networks with reservation services such as ATM or the Internet integrated services (“Intserv”). Applications to other settings are given in the following chapters. We begin the chapter by defining cumulative functions, which can handle both continuous and discrete time models. We show how their use can give a first insight into playout buffer issues, which will be revisited with more detail in Chapter 5. Then the concepts of Leaky Buckets and Generic Cell Rate algorithms are described in the appropriate framework, of arrival curves. We address in detail the most important arrival curves: piecewise linear functions and stair functions. Using the stair functions, we clarify the relation between spacing and arrival curve. We introduce the concept of service curve as a common model for a variety of network nodes. We show that all schedulers generally proposed for ATM or the Internet integrated services can be modeled by a family of simple service curves called the rate-latency service curves. Then we discover physical properties of networks, such as “pay bursts only once” or “greedy shapers keep arrival constraints”. We also discover that greedy shapers are min-plus, time invariant systems. Then we introduce the concept of maximum service curve, which can be used to account for constant delays or for maximum rates. We illustrate all along the chapter how the results can be used for practical buffer dimensioning. We give practical guidelines for handling fixed delays such as propagation delays. We also address the distortions due to variability in packet size.

1.1 1.1.1

M ODELS FOR DATA F LOWS C UMULATIVE ELS

F UNCTIONS , D ISCRETE T IME VERSUS C ONTINUOUS T IME M OD -

It is convenient to describe data flows by means of the cumulative function R(t), defined as the number of bits seen on the flow in time interval [0, t]. By convention, we take R(0) = 0, unless otherwise specified. Function R is always wide-sense increasing, that is, it belongs to the space F defined in Section 3.1.3 on Page 105. We can use a discrete or continuous time model. In real systems, there is always a minimum granularity (bit, word, cell or packet), therefore discrete time with a finite set of values for R(t) could always be assumed. However, it is often computationally simpler to consider continuous time, with a function R that may be continuous or not. If R(t) is a continuous function, we say that we have a fluid model. Otherwise, 3

CHAPTER 1. NETWORK CALCULUS

4

we take the convention that the function is either right or left-continuous (this makes little difference in practice).1 Figure 1.1.1 illustrates these definitions.

C ONVENTION : A flow is described by a wide-sense increasing function R(t); unless otherwise specified, in this book, we consider the following types of models: • discrete time: t ∈ N = {0, 1, 2, 3, ...} • fluid model: t ∈ R+ = [0, +∞) and R is a continuous function • general, continuous time model: t ∈ R+ and R is a left- or right-continuous function

b its

5 k

5 k

R

4 k

1

d (t)

x (t)

3 k

1 k

2

3

4

5

6

7

8

9

R 2*

2 k 1 k

tim e 1

2

3 k

R 1*

2 k

R

4 k

1 0 1 1 1 2 1 3 1 4

1

2

3

4

5

6

7

8

9

1 0 1 1 1 2 1 3 1 4

5 k 4 k

R

3 k

3

2 k 1 k

1

R 2

3

4

5

6

7

8

9

3

*

1 0 1 1 1 2 1 3 1 4

Figure 1.1: Examples of Input and Output functions, illustrating our terminology and convention. R1 and R1∗ show a continuous function of continuous time (fluid model); we assume that packets arrive bit by bit, for a duration of one time unit per packet arrival. R2 and R2∗ show continuous time with discontinuities at packet arrival times (times 1, 4, 8, 8.6 and 14); we assume here that packet arrivals are observed only when the packet has been fully received; the dots represent the value at the point of discontinuity; by convention, we assume that the function is left- or right-continuous. R3 and R3∗ show a discrete time model; the system is observed only at times 0, 1, 2...

t If we assume that R(t) has a derivative dR dt = r(t) such that R(t) = 0 r(s)ds (thus we have a fluid model), then r is called the rate function. Here, however, we will see that it is much simpler to consider cumulative functions such as R rather than rate functions. Contrary to standard algebra, with min-plus algebra we do not need functions to have “nice” properties such as having a derivative. It is always possible to map a continuous time model R(t) to a discrete time model S(n), n ∈ N by choosing a time slot δ and sampling by 1

It would be nice to stick to either left- or right-continuous functions. However, depending on the model, there is no best choice: see Section 1.2.1 and Section 1.7

1.1. MODELS FOR DATA FLOWS

5

S(n) = R(nδ)

(1.1)

In general, this results in a loss of information. For the reverse mapping, we use the following convention. A continuous time model can be derived from S(n), n ∈ N by letting2 t R (t) = S( ) δ

(1.2)

The resulting function R is always left-continuous, as we already required. Figure 1.1.1 illustrates this mapping with δ = 1, S = R3 and R = R2 . Thanks to the mapping in Equation (1.1), any result for a continuous time model also applies to discrete time. Unless otherwise stated, all results in this book apply to both continuous and discrete time. Discrete time models are generally used in the context of ATM; in contrast, handling variable size packets is usually done with a continuous time model (not necessarily fluid). Note that handling variable size packets requires some specific mechanisms, described in Section 1.7. Consider now a system S, which we view as a blackbox; S receives input data, described by its cumulative function R(t), and delivers the data after a variable delay. Call R∗ (t) the output function, namely, the cumulative function at the output of system S. System S might be, for example, a single buffer served at a constant rate, a complex communication node, or even a complete network. Figure 1.1.1 shows input and output functions for a single server queue, where every packet takes exactly 3 time units to be served. With output function R1∗ (fluid model) the assumption is that a packet can be served as soon as a first bit has arrived (cut-through assumption), and that a packet departure can be observed bit by bit, at a constant rate. For example, the first packet arrives between times 1 and 2, and leaves between times 1 and 4. With output function R2∗ the assumption is that a packet is served as soon as it has been fully received and is considered out of the system only when it is fully transmitted (store and forward assumption). Here, the first packet arrives immediately after time 1, and leaves immediately after time 4. With output function R3∗ (discrete time model), the first packet arrives at time 2 and leaves at time 5.

1.1.2

BACKLOG AND V IRTUAL D ELAY

From the input and output functions, we derive the two following quantities of interest. D EFINITION 1.1.1 (BACKLOG AND D ELAY ). For a lossless system: • The backlog at time t is R(t) − R∗ (t). • The virtual delay at time t is d(t) = inf {τ ≥ 0 : R(t) ≤ R∗ (t + τ )} The backlog is the amount of bits that are held inside the system; if the system is a single buffer, it is the queue length. In contrast, if the system is more complex, then the backlog is the number of bits “in transit”, assuming that we can observe input and output simultaneously. The virtual delay at time t is the delay that would be experienced by a bit arriving at time t if all bits received before it are served before it. In Figure 1.1.1, the backlog, called x(t), is shown as the vertical deviation between input and output functions. The virtual delay is the horizontal deviation. If the input and output function are continuous (fluid model), then it is easy to see that R∗ (t + d(t)) = R(t), and that d(t) is the smallest value satisfying this equation. In Figure 1.1.1, we see that the values of backlog and virtual delay slightly differ for the three models. Thus the delay experienced by the last bit of the first packet is d(2) = 2 time units for the first subfigure; in contrast, it is equal to d(1) = 3 time units on the second subfigure. This is of course in accordance with the 2

x (“ceiling of x”) is defined as the smallest integer ≥ x; for example 2.3 = 3 and 2 = 2


6

different assumptions made for each of the models. Similarly, the delay for the fourth packet on subfigure 2 is d(8.6) = 5.4 time units, which corresponds to 2.4 units of waiting time and 3 units of service time. In contrast, on the third subfigure, it is equal to d(9) = 6 units; the difference is the loss of accuracy resulting from discretization.

1.1.3

E XAMPLE : T HE P LAYOUT B UFFER

Cumulative functions are a powerful tool for studying delays and buffers. In order to illustrate this, consider the simple playout buffer problem that we describe now. Consider a packet switched network that carries bits of information from a source with a constant bit rate r (Figure 1.2) as is the case for example, with circuit emulation. We take a fluid model, as illustrated in Figure 1.2. We have a first system S, the network, with input function R(t) = rt. The network imposes some variable delay, because of queuing points, therefore the output R∗ does not have a constant rate r. What can be done to recreate a constant bit stream ? A standard mechanism is to smooth the delay variation in a playout buffer. It operates as follows. When b it s

R (t )

R * (t )

S (t ) (D 2 )

- d (0

+) +

D )

S (t ) S

R * (t )

(t 2 ) : r (D

1) :

r( t

S

R * (t )

- d (0

R (t )

+) -

D )

d (t )

t im e

t 1

(D

d (0 ) - D d (0 + ) d (0 + ) + D

Figure 1.2: A Simple Playout Buffer Example the first bit of data arrives, at time dr (0), where dr (0) = limt→0,t>0 d(t) is the limit to the right of function d3 , it is stored in the buffer until a fixed time ∆ has elapsed. Then the buffer is served at a constant rate r whenever it is not empty. This gives us a second system S , with input R∗ and output S. Let us assume that the network delay variation is bounded by ∆. This implies that for every time t, the virtual delay (which is the real delay in that case) satisfies −∆ ≤ d(t) − dr (0) ≤ ∆ Thus, since we have a fluid model, we have r(t − dr (0) − ∆) ≤ R∗ (t) ≤ r(t − dr (0) + ∆) which is illustrated in the figure by the two lines (D1) and (D2) parallel to R(t). The figure suggests that, for the playout buffer S the input function R∗ is always above the straight line (D2), which means that the playout buffer never underflows. This suggests in turn that the output function S(t) is given by S(t) = r(t − dr (0) − ∆). Formally, the proof is as follows. We proceed by contradiction. Assume the buffer starves at some time, and let t1 be the first time at which this happens. Clearly the playout buffer is empty at time t1 , thus R∗ (t1 ) = S(t1 ). There is a time interval [t1 , t1 + ] during which the number of bits arriving at the playout buffer is less than r (see Figure 1.2). Thus, d(t1 + ) > dr (0) + ∆ which is not possible. Secondly, the 3

It is the virtual delay for a hypothetical bit that would arrive just after time 0. Other authors often use the notation d(0+)

1.2. ARRIVAL CURVES

7

backlog in the buffer at time t is equal to R∗ (t) − S(t), which is bounded by the vertical deviation between (D1) and (D2), namely, 2r∆. We have thus shown that the playout buffer is able to remove the delay variation imposed by the network. We summarize this as follows. P ROPOSITION 1.1.1. Consider a constant bit rate stream of rate r, modified by a network that imposes a variable delay variation and no loss. The resulting flow is put into a playout buffer, which operates by delaying the first bit of the flow by ∆, and reading the flow at rate r. Assume that the delay variation imposed by the network is bounded by ∆, then 1. the playout buffer never starves and produces a constant output at rate r; 2. a buffer size of 2∆r is sufficient to avoid overflow. We study playout buffers in more details in Chapter 5, using the network calculus concepts further introduced in this chapter.

1.2 1.2.1

A RRIVAL C URVES D EFINITION OF AN A RRIVAL C URVE

Assume that we want to provide guarantees to data flows. This requires some specific support in the network, as explained in Section 1.3; as a counterpart, we need to limit the traffic sent by sources. With integrated services networks (ATM or the integrated services internet), this is done by using the concept of arrival curve, defined below. D EFINITION 1.2.1 (A RRIVAL C URVE ). Given a wide-sense increasing function α defined for t ≥ 0 we say that a flow R is constrained by α if and only if for all s ≤ t: R(t) − R(s) ≤ α(t − s) We say that R has α as an arrival curve, or also that R is α-smooth. Note that the condition is over a set of overlapping intervals, as Figure 1.3 illustrates. b its t - s

b its

R (t)

a (t)

s

tim e tim e t

Figure 1.3: Example of Constraint by arrival curve, showing a cumulative function R(t) constrained by the arrival curve α(t).


8

A FFINE A RRIVAL C URVES : For example, if α(t) = rt, then the constraint means that, on any time window of width τ , the number of bits for the flow is limited by rτ . We say in that case that the flow is peak rate limited. This occurs if we know that the flow is arriving on a link whose physical bit rate is limited by r b/s. A flow where the only constraint is a limit on the peak rate is often (improperly) called a “constant bit rate” (CBR) flow, or “deterministic bit rate” (DBR) flow. Having α(t) = b, with b a constant, as an arrival curve means that the maximum number of bits that may ever be sent on the flow is at most b. More generally, because of their relationship with leaky buckets, we will often use affine arrival curves γr,b , defined by: γr,b (t) = rt+b for t > 0 and 0 otherwise. Having γr,b as an arrival curve allows a source to send b bits at once, but not more than r b/s over the long run. Parameters b and r are called the burst tolerance (in units of data) and the rate (in units of data per time unit). Figure 1.3 illustrates such a constraint. S TAIR F UNCTIONS AS A RRIVAL C URVES : In the context of ATM, we also use arrival curves of the form kvT,τ , where vT,τ is the stair functions defined by vT,τ (t) = t+τ T for t > 0 and 0 otherwise (see Section 3.1.3 for an illustration). Note that vT,τ (t) = vT,0 (t + τ ), thus vT,τ results from vT,0 by a time shift to the left. Parameter T (the “interval”) and τ (the “tolerance”) are expressed in time units. In order to understand the use of vT,τ , consider a flow that sends packets of a fixed size, equal to k unit of data (for example, an ATM flow). Assume that the packets are spaced by at least T time units. An example is a constant bit rate voice encoder, which generates packets periodically during talk spurts, and is silent otherwise. Such a flow has kvT,0 as an arrival curve. Assume now that the flow is multiplexed with some others. A simple way to think of this scenario is to assume that the packets are put into a queue, together with other flows. This is typically what occurs at a workstation, in the operating system or at the ATM adapter. The queue imposes a variable delay; assume it can be bounded by some value equal to τ time units. We will see in the rest of this chapter and in Chapter 2 how we can provide such bounds. Call R(t) the input function for the flow at the multiplexer, and R∗ (t) the output function. We have R∗ (s) ≤ R(s − τ ), from which we derive: R∗ (t) − R∗ (s) ≤ R(t) − R(s − τ ) ≤ kvT,0 (t − s + τ ) = kvT,τ (t − s) Thus R∗ has kvT,τ as an arrival curve. We have shown that a periodic flow, with period T , and packets of constant size k, that suffers a variable delay ≤ τ , has kvT,τ as an arrival curve. The parameter τ is often called the “one-point cell delay variation”, as it corresponds to a deviation from a periodic flow that can be observed at one point. In general, function vT,τ can be used to express minimum spacing between packets, as the following proposition shows. P ROPOSITION 1.2.1 (S PACING AS AN ARRIVAL CONSTRAINT ). Consider a flow, with cumulative function R(t), that generates packets of constant size equal to k data units, with instantaneous packet arrivals. Assume time is discrete or time is continuous and R is left-continuous. Call tn the arrival time for the nth packet. The following two properties are equivalent: 1. for all m, n, tm+n − tm ≥ nT − τ 2. the flow has kvT,τ as an arrival curve The conditions on packet size and packet generation mean that R(t) has the form nk, with n ∈ N. The spacing condition implies that the time interval between two consecutive packets is ≥ T − τ , between a packet and the next but one is ≥ 2T − τ , etc. P ROOF : Assume that property 1 holds. Consider an arbitrary interval ]s, t], and call n the number of packet arrivals in the interval. Say that these packets are numbered m + 1, . . . , m + n, so that s < tm+1 ≤

1.2. ARRIVAL CURVES

9

. . . ≤ tm+n ≤ t, from which we have t − s > tm+n − tm+1 Combining with property 1, we get t − s > (n − 1)T − τ From the definition of vT,τ it follows that vT,τ (t − s) ≥ n. Thus R(t) − R(s) ≤ kvT,τ (t − s), which shows the first part of the proof. Conversely, assume now that property 2 holds. If time is discrete, we convert the model to continuous time using the mapping in Equation 1.2, thus we can consider that we are in the continuous time case. Consider some arbitrary integers m, n; for all > 0, we have, under the assumption in the proposition: R(tm+n + ) − R(tm ) ≥ (n + 1)k thus, from the definition of vT,τ , tm+n − tm + > nT − τ This is true for all > 0, thus tm+n − tm ≥ nT − τ . In the rest of this section we clarify the relationship between arrival curve constraints defined by affine and by stair functions. First we need a technical lemma, which amounts to saying that we can always change an arrival curve to be left-continuous. L EMMA 1.2.1 (R EDUCTION TO LEFT- CONTINUOUS ARRIVAL CURVES ). Consider a flow R(t) and a wide sense increasing function α(t), defined for t ≥ 0. Assume that R is either left-continuous, or rightcontinuous. Denote with αl (t) the limit to the left of α at t (this limit exists at every point because α is wide sense increasing); we have αl (t) = sups 1. This flow might result from the superposition of several ATM flows. You can convince yourself that this constraint cannot be mapped to a constraint of the form γr,b . We will come back to this example in Section 1.4.1.

1.2.2

L EAKY B UCKET AND G ENERIC C ELL R ATE A LGORITHM

Arrival curve constraints find their origins in the concept of leaky bucket and generic cell rate algorithms, which we describe now. We show that leaky buckets correspond to affine arrival curves γr,b , while the generic cell rate algorithm corresponds to stair functions vT,τ . For flows of fixed size packets, such as ATM cells, the two are thus equivalent. D EFINITION 1.2.2 (L EAKY B UCKET C ONTROLLER ). A Leaky Bucket Controller is a device that analyzes the data on a flow R(t) as follows. There is a pool (bucket) of fluid of size b. The bucket is initially empty. The bucket has a hole and leaks at a rate of r units of fluid per second when it is not empty. Data from the flow R(t) has to pour into the bucket an amount of fluid equal to the amount of data. Data that would cause the bucket to overflow is declared non-conformant, otherwise the data is declared conformant. Figure 1.2.2 illustrates the definition. Fluid in the leaky bucket does not represent data, however, it is counted in the same unit as data. Data that is not able to pour fluid into the bucket is said to be “non-conformant” data. In ATM systems, non-conformant data is either discarded, tagged with a low priority for loss (“red” cells), or can be put in a buffer (buffered leaky bucket controller). With the Integrated Services Internet, non-conformant data is in principle not marked, but simply passed as best effort traffic (namely, normal IP traffic). 5 k

4 J

b its

R (t)

4 k 3 k

4 J >

2 k

N J

H

N J

1 k

1

2

3

4

5

6

7

8

9

1 0 1 1 1 2 1 3 1 4

Figure 1.4: A Leaky Bucket Controller. The second part of the figure shows (in grey) the level of the bucket x(t) for a sample input, with r = 0.4 kbits per time unit and b = 1.5 kbits. The packet arriving at time t = 8.6 is not conformant, and no fluid is added to the bucket. If b would be equal to 2 kbits, then all packets would be conformant.

1.2. ARRIVAL CURVES

11

We want now to show that a leaky bucket controller enforces an arrival curve constraint equal to γr,b . We need the following lemma. L EMMA 1.2.2. Consider a buffer served at a constant rate r. Assume that the buffer is empty at time 0. The input is described by the cumulative function R(t). If there is no overflow during [0, t], the buffer content at time t is given by x(t) = sup {R(t) − R(s) − r(t − s)} s:s≤t

P ROOF : The lemma can be obtained as a special case of Corollary 1.5.2 on page 32, however we give here a direct proof. First note that for all s such that s ≤ t, (t − s)r is an upper bound on the number of bits output in ]s, t], therefore: R(t) − R(s) − x(t) + x(s) ≤ (t − s)r Thus x(t) ≥ R(t) − R(s) + x(s) − (t − s)r ≥ R(t) − R(s) − (t − s)r which proves that x(t) ≥ sups:s≤t {R(t) − R(s) − r(t − s)}. Conversely, call t0 the latest time at which the buffer was empty before time t: t0 = sup{s : s ≤ t, x(s) = 0} (If x(t) > 0 then t0 is the beginning of the busy period at time t). During ]t0 , t], the queue is never empty, therefore it outputs bit at rate r, and thus x(t) = x(t0 ) + R(t) − R(t0 ) − (t − t0 )r

(1.3)

We assume that R is left-continuous (otherwise the proof is a little more complex); thus x(t0 ) = 0 and thus x(t) ≤ sups:s≤t {R(t) − R(s) − r(t − s)} Now the content of a leaky bucket behaves exactly like a buffer served at rate r, and with capacity b. Thus, a flow R(t) is conformant if and only if the bucket content x(t) never exceeds b. From Lemma 1.2.2, this means that sup {R(t) − R(s) − r(t − s)} ≤ b s:s≤t

which is equivalent to R(t) − R(s) ≤ r(t − s) + b for all s ≤ t. We have thus shown the following. P ROPOSITION 1.2.3. A leaky bucket controller with leak rate r and bucket size b forces a flow to be constrained by the arrival curve γr,b , namely: 1. the flow of conformant data has γr,b as an arrival curve; 2. if the input already has γr,b as an arrival curve, then all data is conformant. We will see in Section 1.4.1 a simple interpretation of the leaky bucket parameters, namely: r is the minimum rate required to serve the flow, and b is the buffer required to serve the flow at a constant rate. Parallel to the concept of leaky bucket is the Generic Cell Rate Algorithm (GCRA), used with ATM. D EFINITION 1.2.3 (GCRA (T, τ )). The Generic Cell Rate Algorithm (GCRA) with parameters (T, τ ) is used with fixed size packets, called cells, and defines conformant cells as follows. It takes as input a cell arrival time t and returns result. It has an internal (static) variable tat (theoretical arrival time).


12

• initially, tat = 0 • when a cell arrives at time t, then if (t < tat - tau) result = NON-CONFORMANT; else { tat = max (t, tat) + T; result = CONFORMANT; } Table 1.1 illustrate the definition of GCRA. It illustrates that T1 is the long term rate that can be sustained by the flow (in cells per time unit); while τ is a tolerance that quantifies how early cells may arrive with respect to an ideal spacing of T between cells. We see on the first example that cells may be early by 2 time units (cells arriving at times 18 to 48), however this may not be cumultated, otherwise the rate of T1 would be exceeded (cell arriving at time 57). arrival time tat before arrival result

0 0 c

arrival time tat before arrival result

10 10 c

18 20 c

28 30 c

38 40 c

0 0 c

10 10 c

15 20 non-c

48 50 c 25 20 c

57 60 non-c 35 30 c

Table 1.1: Examples for GCRA(10,2). The table gives the cell arrival times, the value of the tat internal variable just before the cell arrival, and the result for the cell (c = conformant, non-c = non-conformant). In general, we have the following result, which establishes the relationship between GCRA and the stair functions vT,τ . P ROPOSITION 1.2.4. Consider a flow, with cumulative function R(t), that generates packets of constant size equal to k data units, with instantaneous packet arrivals. Assume time is discrete or time is continuous and R is left-continuous. The following two properties are equivalent: 1. the flow is conformant to GCRA(T, τ ) 2. the flow has (k vT,τ ) as an arrival curve P ROOF : The proof uses max-plus algebra. Assume that property 1 holds. Denote with θn the value of tat just after the arrival of the nth packet (or cell), and by convention θ0 = 0. Also call tn the arrival time of the nth packet. From the definition of the GCRA we have θn = max(tn , θn−1 ) + T . We write this equation for all m ≤ n, using the notation ∨ for max. The distributivity of addition with respect to ∨ gives: ⎧ θn = (θn−1 + T ) ∨ (tn + T ) ⎪ ⎪ ⎨ θn−1 + T = (θn−2 + 2T ) ∨ (tn−1 + 2T ) . .. ⎪ ⎪ ⎩ θ1 + (n − 1)T = (θ0 + nT ) ∨ (t1 + nT ) Note that (θ0 + nT ) ∨ (t1 + nT ) = t1 + nT because θ0 = 0 and t1 ≥ 0, thus the last equation can be simplified to θ1 + (n − 1)T = t1 + nT . Now the iterative substitution of one equation into the previous one, starting from the last one, gives θn = (tn + T ) ∨ (tn−1 + 2T ) ∨ . . . ∨ (t1 + nT )

(1.4)

1.2. ARRIVAL CURVES

13

Now consider the (m + n)th arrival, for some m, n ∈ N, with m ≥ 1. By property 1, the packet is conformant, thus (1.5) tm+n ≥ θm+n−1 − τ Now from Equation (1.4), θm+n−1 ≥ tj + (m + n − j)T for all 1 ≤ j ≤ m + n − 1. For j = m, we obtain θm+n−1 ≥ tm + nT . Combining this with Equation (1.5), we have tm+n ≥ tm + nT − τ . With proposition 1.2.1, this shows property 2. Conversely, assume now that property 2 holds. We show by induction on n that the nth packet is conformant. This is always true for n = 1. Assume it is true for all m ≤ n. Then, with the same reasoning as above, Equation (1.4) holds for n. We rewrite it as θn = max1≤j≤n {tj +(n−j+1)T }. Now from proposition 1.2.1, tn+1 ≥ tj + (n − j + 1)T − τ for all 1 ≤ j ≤ n, thus tn+1 ≥ max1≤j≤n {tj + (n − j + 1)T } − τ . Combining the two, we find that tn+1 ≥ θn − τ , thus the (n + 1)th packet is conformant. Note the analogy between Equation (1.4) and Lemma 1.2.2. Indeed, from proposition 1.2.2, for packets of constant size, there is equivalence between arrival constraints by affine functions γr,b and by stair functions vT,τ . This shows the following result. C OROLLARY 1.2.1. For a flow with packets of constant size, satisfying the GCRA(T, τ ) is equivalent to satisfying a leaky bucket controller, with rate r and burst tolerance b given by: b=(

τ + 1)δ T

r=

δ T

In the formulas, δ is the packet size in units of data. The corollary can also be shown by a direct equivalence of the GCRA algorithm to a leaky bucket controller. Take the ATM cell as unit of data. The results above show that for an ATM cell flow, being conformant to GCRA(T, τ ) is equivalent to having vT,τ as an arrival curve. It is also equivalent to having γr,b as an arrival curve, with r = T1 and b = Tτ + 1. Consider a family of I leaky bucket controllers (or GCRAs), with parameters ri , bi , for 1 ≤ i ≤ I. If we apply all of them in parallel to the same flow, then the conformant data is data that is conformant for each of the controllers in isolation. The flow of conformant data has as an arrival curve α(t) = min (γri ,bi (t)) = min (ri t + bi ) 1≤i≤I

1≤i≤I

It can easily be shown that the family of arrival curves that can be obtained in this way is the set of concave, piecewise linear functions, with a finite number of pieces. We will see in Section 1.5 some examples of functions that do not belong to this family. A PPLICATION TO ATM AND THE I NTERNET Leaky buckets and GCRA are used by standard bodies to define conformant flows in Integrated Services Networks. With ATM, a constant bit rate connection (CBR) is defined by one GCRA (or equivalently, one leaky bucket), with parameters (T, τ ). T is called the ideal cell interval, and τ is called the Cell Delay Variation Tolerance (CDVT). Still with ATM, a variable bit rate (VBR) connection is defined as one connection with an arrival curve that corresponds to 2 leaky buckets or GCRA controllers. The Integrated services framework of the Internet (Intserv) uses the same family of arrival curves, such as α(t) = min(M + pt, rt + b) (1.6) where M is interpreted as the maximum packet size, p as the peak rate, b as the burst tolearance, and r as the sustainable rate (Figure 1.5). In Intserv jargon, the 4-uple (p, M, r, b) is also called a T-SPEC (traffic specification).


14

ra te p ra te r b M

Figure 1.5: Arrival curve for ATM VBR and for Intserv flows

1.2.3

S UB - ADDITIVITY AND A RRIVAL C URVES

In this Section we discover the fundamental relationship between min-plus algebra and arrival curves. Let us start with a motivating example. Consider a flow R(t) ∈ N with t ∈ N; for example the flow is an ATM cell flow, counted in cells. Time is discrete to simplify the discussion. Assume that we know that the flow is constrained by the arrival curve 3v10,0 ; for example, the flow is the superposition of 3 CBR connections of peak rate 0.1 cell per time unit each. Assume in addition that we know that the flow arrives at the point of observation over a link with a physical characteristic of 1 cell per time unit. We can conclude that the flow is also constrained by the arrival curve v1,0 . Thus, obviously, it is constrained by α1 = min(3v10,0 , v1,0 ). Figure 1.6 shows the function α1 . c e lls

c e lls

6

6 3

3 1 0

tim e s lo ts

1 0

tim e s lo ts

Figure 1.6: The arrival curve α1 = min(3v10,0 , v1,0 ) on the left, and its sub-additive closure (“good” function) α¯1 on the right. Time is discrete, lines are put for ease of reading. Now the arrival curve α1 tells us that R(10) ≤ 3 and R(11) ≤ 6. However, since there can arrive at most 1 cell per time unit , we can also conclude that R(11) ≤ R(10) + [R(11) − R(10)] ≤ α1 (10) + α1 (1) = 4. In other words, the sheer knowledge that R is constrained by α1 allows us to derive a better bound than α1 itself. This is because α1 is not a “good” function, in a sense that we define now. D EFINITION 1.2.4. Consider a function α in F. We say that α is a “good” function if any one of the following equivalent properties is satisfied 1. 2. 3. 4.

α is sub-additive and α(0) = 0 α=α⊗α αα=α α=α ¯ (sub-additive closure of α).

The definition uses the concepts of sub-additivity, min-plus convolution, min-plus deconvolution and subadditive closure, which are defined in Chapter 3. The equivalence between the four items comes from Corollaries 3.1.1 on page 120 and 3.1.13 on page 125. Sub-additivity (item 1) means that α(s + t) ≤

1.2. ARRIVAL CURVES

15

α(s) + α(t). If α is not sub-additive, then α(s) + α(t) may be a better bound than α(s + t), as is the case with α1 in the example above. Item 2, 3 and 4 use the concepts of min-plus convolution, min-plus deconvolution and sub-additive closure, defined in Chapter 3. We know in particular (Theorem 3.1.10) that the sub-additive closure of a function α is the largest “good” function α ¯ such that α ¯ ≤ α. We also know that α ¯ ∈ F if α ∈ F. The main result about arrival curves is that any arrival curve can be replaced by its sub-additive closure, which is a “good” arrival curve. Figure 1.6 shows α¯1 for our example above. T HEOREM 1.2.1 (R EDUCTION OF A RRIVAL C URVE TO A S UB -A DDITIVE O NE ). Saying that a flow is constrained by a wide-sense increasing function α is equivalent to saying that it is constrained by the subadditive closure α ¯. The proof of the theorem leads us to the heart of the concept of arrival curve, namely, its correspondence with a fundamental, linear relationships in min-plus algebra, which we will now derive. L EMMA 1.2.3. A flow R is constrained by arrival curve α if and only if R ≤ R ⊗ α P ROOF : Remember that an equation such as R ≤ R ⊗ α means that for all times t, R(t) ≤ (R ⊗ α)(t). The min-plus convolution R ⊗ α is defined in Chapter 3, page 111; since R(s) and α(s) are defined only for s ≥ 0, the definition of R ⊗ α is: (R ⊗ α)(t) = inf 0≤s≤t (R(s) + α(t − s)). Thus R ≤ R ⊗ α is equivalent to R(t) ≤ R(s) + α(t − s) for all 0 ≤ s ≤ t. L EMMA 1.2.4. If α1 and α2 are arrival curves for a flow R, then so is α1 ⊗ α2 P ROOF : We know from Chapter 3 that α1 ⊗ α2 is wide-sense increasing if α1 and α2 are. The rest of the proof follows immediately from Lemma 1.2.3 and the associativity of ⊗. P ROOF OF T HEOREM Since α is an arrival curve, so is α ⊗ α, and by iteration, so is α(n) for all n ≥ 1. By the definition of δ0 , it is also an arrival curve. Thus so is α ¯ = inf n≥0 α(n) . Conversely, α ¯ ≤ α; thus, if α ¯ is an arrival curve, then so is α. E XAMPLES We should thus restrict our choice of arrival curves to sub-additive functions. As we can expect, the functions γr,b and vT,τ introduced in Section 1.2.1 are sub-additive and since their value is 0 for t = 0, they are “good” functions, as we now show. Indeed, we know from Chapter 1 that any concave function α such that α(0) = 0 is sub-additive. This explains why the functions γr,b are sub-additive. Functions vT,τ are not concave, but they still are sub-additive. This is because, from its very definition, the ceiling function is sub-additive, thus vT,τ (s + t) =

s+τ t s+τ t+τ s+t+τ

≤

+ ≤

+

= vT,τ (s) + vT,τ (t) T T T T T

Let us return to our introductory example with α1 = min(3v10,0 , v1,0 ). As we discussed, α1 is not subadditive. From Theorem 1.2.1, we should thus replace α1 by its sub-additive closure α¯1 , which can be computed by Equation (3.13). The computation is simplified by the following remark, which follows immediately from Theorem 3.1.11: L EMMA 1.2.5. Let γ1 and γ2 be two “good” functions. The sub-additive closure of min(γ1 , γ2 ) is γ1 ⊗ γ2 . We can apply the lemma to α1 = 3v10,0 ∧ v1,0 , since vT,τ is a “good” function. Thus α¯1 = 3v10,0 ⊗ v1,0 , which the alert reader will enjoy computing. The result is plotted in Figure 1.6. Finally, let us mention the following equivalence, the proof of which is easy and left to the reader.


16

P ROPOSITION 1.2.5. For a given wide-sense increasing function α, with α(0) = 0, consider a source defined by R(t) = α(t) (greedy source). The source has α as an arrival curve if and only if α is a “good” function. VBR ARRIVAL CURVE Now let us examine the family of arrival curves obtained by combinations of leaky buckets or GCRAs (concave piecewise linear functions). We know from Chapter 3 that if γ1 and γ2 are concave, with γ1 (0) = γ2 (0) = 0, then γ1 ⊗ γ2 = γ1 ∧ γ2 . Thus any concave piecewise linear function α such that α(0) = 0 is a “good” function. In particular, if we define the arrival curve for VBR connections or Intserv flows by α(t) = min(pt + M, rt + b) if t > 0 α(0) = 0 (see Figure 1.5) then α is a “good” function. We have seen in Lemma 1.2.1 that an arrival curve α can always be replaced by its limit to the left αl . We might wonder how this combines with the sub-additive closure, and in particular, whether these two operations commute (in other words, do we have (¯ α)l = αl ?). In general, if α is left-continuous, then we cannot guarantee that α ¯ is also left-continuous, thus we cannot guarantee that the operations commute. α)l = (¯ α)l . Starting from an arrival However, it can be shown that (¯ α)l is always a “good” function, thus (¯ curve α we can therefore improve by taking the sub-additive closure first, then the limit to the left. The resulting arrival curve (¯ α)l is a “good” function that is also left-continuous (a “very good” function), and the constraint by α is equivalent to the constraint by (¯ α)l Lastly, let us mention that it can easily be shown, using an argument of uniform continuity, that if α takes only a finite set of values over any bounded time interval, and if α is left-continuous, then so is α ¯ and then we do have (¯ α)l = αl . This assumption is always true in discrete time, and in most cases in practice.

1.2.4

M INIMUM A RRIVAL C URVE

Consider now a given flow R(t), for which we would like to determine a minimal arrival curve. This problem arises, for example, when R is known from measurements. The following theorem says that there is indeed one minimal arrival curve. T HEOREM 1.2.2 (M INIMUM A RRIVAL C URVE ). Consider a flow R(t)t≥0 . Then • function R R is an arrival curve for the flow • for any arrival curve α that constrains the flow, we have: (R R) ≤ α • R R is a “good” function Function R R is called the minimum arrival curve for flow R. The minimum arrival curve uses min-plus deconvolution, defined in Chapter 3. Figure 1.2.4 shows an example of R R for a measured function R. P ROOF : By definition of , we have (R R)(t) = supv≥0 {R(t + v) − R(v)}, it follows that (R R) is an arrival curve. Now assume that some α is also an arrival curve for R. From Lemma 1.2.3, we have R ≤ R ⊗ α). From Rule 14 in Theorem 3.1.12 in Chapter 3, it follows that R R ≤ α, which shows that R R is the minimal arrival curve for R. Lastly, R R is a “good” function from Rule 15 in Theorem 3.1.12. Consider a greedy source, with R(t) = α(t), where α is a “good” function. What is the minimum arrival curve ?5 Lastly, the curious reader might wonder whether R R is left-continuous. The answer is as 5

Answer: from the equivalence in Definition 1.2.4, the minimum arrival curve is α itself.

1.2. ARRIVAL CURVES

17

70 60 50 40 30 20 10 100

200

300

400

100

200

300

400

70 60 50 40 30 20 10

10000 8000 6000 4000 2000 100

200

300

400

Figure 1.7: Example of minimum arrival curve. Time is discrete, one time unit is 40 ms. The top figures shows, for two similar traces, the number of packet arrivals at every time slot. Every packet is of constant size (416 bytes). The bottom figure shows the minimum arrival curve for the first trace (top curve) and the second trace (bottom curve). The large burst in the first trace comes earlier, therefore its minimum arrival curve is slightly larger.


18

follows. Assume that R is either right or left-continuous. By lemma 1.2.1, the limit to the left (R R)l is also an arrival curve, and is bounded from above by R R. Since R R is the minimum arrival curve, it follows that (R R)l = R R, thus R R is left-continuous (and is thus a “very good” function). In many cases, one is interested not in the absolute minimum arrival curve as presented here, but in a minimum arrival curve within a family of arrival curves, for example, among all γr,b functions. For a development along this line, see [61].

1.3 1.3.1

S ERVICE C URVES D EFINITION OF S ERVICE C URVE

We have seen that one first principle in integrated services networks is to put arrival curve constraints on flows. In order to provide reservations, network nodes in return need to offer some guarantees to flows. This is done by packet schedulers [45]. The details of packet scheduling are abstracted using the concept of service curve, which we introduce and study in this section. Since the concept of service curve is more abstract than that of arrival curve, we introduce it on some examples. A first, simple example of a scheduler is a Generalized Processor Sharing (GPS) node [63]. We define now a simple view of GPS; more details are given in Chapter 2. A GPS node serves several flows in parallel, and we can consider that every flow is allocated a given rate. The guarantee is that during a period of duration t, for which a flow has some backlog in the node, it receives an amount of service at least equal to rt, where r is its allocated rate. A GPS node is a theoretical concept, which is not really implementable, because it relies on a fluid model, while real networks use packets. We will see in Section 2.1 on page 67 how to account for the difference between a real implementation and GPS. Consider a input flow R, with output R∗ , that is served in a GPS node, with allocated rate r. Let us also assume that the node buffer is large enough so that overflow is not possible. We will see in this section how to compute the buffer size required to satisfy this assumption. Lossy systems are the object of Chapter 9. Under these assumptions, for all time t, call t0 the beginning of the last busy period for the flow up to time t. From the GPS assumption, we have R∗ (t) − R∗ (t0 ) ≥ r(t − t0 ) Assume as usual that R is left-continuous; at time t0 the backlog for the flow is 0, which is expressed by R(t0 ) − R∗ (t0 ) = 0. Combining this with the previous equation, we obtain: R∗ (t) − R(t0 ) ≥ r(t − t0 ) We have thus shown that, for all time t: R∗ (t) ≥ inf 0≤s≤t [R(s) + r(t − s)], which can be written as R∗ ≥ R ⊗ γr,0

(1.7)

Note that a limiting case of GPS node is the constant bit rate server with rate r, dedicated to serving a single flow. We will study GPS in more details in Chapter 2. Consider now a second example. Assume that the only information we have about a network node is that the maximum delay for the bits of a given flow R is bounded by some fixed value T , and that the bits of the flow are served in first in, first out order. We will see in Section 1.5 that this is used with a family of schedulers called “earliest deadline first” (EDF). We can translate the assumption on the delay bound to d(t) ≤ T for all t. Now since R∗ is always wide-sense increasing, it follows from the definition of d(t) that R∗ (t + T ) ≥ R(t). Conversely, if R∗ (t + T ) ≥ R(t), then d(t) ≤ T . In other words, our condition that the maximum delay is bounded by T is equivalent to R∗ (t + T ) ≥ R(t) for all t. This in turn can be re-written as R∗ (s) ≥ R(s − T )

1.3. SERVICE CURVES

19

for all s ≥ T . We have introduced in Chapter 3 the “impulse” function δT defined by δT (t) = 0 if 0 ≤ t ≤ T and δT (t) = +∞ if t > T . It has the property that, for any wide-sense increasing function x(t), defined for t ≤ 0, (x ⊗ δT )(t) = x(t − T ) if t ≥ T and (x ⊗ δT )(t) = x(0) otherwise. Our condition on the maximum delay can thus be written as (1.8) R ∗ ≥ R ⊗ δT For the two examples above, there is an input-output relationship of the same form (Equations (1.7) and (1.8)). This suggests the definition of service curve, which, as we see in the rest of this section, is indeed able to provide useful results.

R * (t)

R (t)

d a ta

b (t) tim e

(R Ä

b )(t)

tim e

Figure 1.8: Definition of service curve. The output R∗ must be above R ⊗ β, which is the lower envelope of all curves

t → R(t0 ) + β(t − t0 ).

D EFINITION 1.3.1 (S ERVICE C URVE ). Consider a system S and a flow through S with input and output function R and R∗ . We say that S offers to the flow a service curve β if and only if β is wide sense increasing, β(0) = 0 and R∗ ≥ R ⊗ β Figure 1.8 illustrates the definition. The definition means that β is a wide sense increasing function, with β(0) = 0, and that for all t ≥ 0, R∗ (t) ≥ inf (R(s) + β(t − s)) s≤t

In practice, we can avoid the use of an infimum if β is continuous. The following proposition is an immediate consequence of Theorem 3.1.8 on Page 115. P ROPOSITION 1.3.1. If β is continuous, the service curve property means that for all t we can find t0 ≤ t such that (1.9) R∗ (t) ≥ Rl (t0 ) + β(t − t0 ) where Rl (t0 ) = sup{s R∗ (t + τ ). Now the service curve property at time t + τ implies that there is some s0 such that R(t) > R(t + τ − s0 ) + β(s0 ) It follows from this latter equation that t + τ − s0 < t. Thus α(τ − s0 ) ≥ [R(t) − R(t + τ − s0 )] > β(s0 ) Thus τ ≤ δ(τ − s0 ) ≤ h(α, β). This is true for all τ < d(t) thus d(t) ≤ h(α, β). T HEOREM 1.4.3 (O UTPUT F LOW ). Assume a flow, constrained by arrival curve α, traverses a system that offers a service curve of β. The output flow is constrained by the arrival curve α∗ = α β. The theorem uses min-plus deconvolution, introduced in Chapter 3, which we have already used in Theorem 1.2.2. P ROOF : With the same notation as above, consider R∗ (t) − R∗ (t − s), for 0 ≤ t − s ≤ t. Consider the definition of the service curve, applied at time t − s. Assume for a second that the inf in the definition of R ⊗ β is a min, that is to say, there is some u ≥ 0 such that 0 ≤ t − s − u and (R ⊗ β)(t − s) = R(t − s − u) + β(u) Thus and thus

R∗ (t − s) − R(t − s − u) ≥ β(u) R∗ (t) − R∗ (t − s) ≤ R∗ (t) − β(u) − R(t − s − u)

Now R∗ (t) ≤ R(t), therefore R∗ (t) − R∗ (t − s) ≤ R(t) − R(t − s − u) − β(u) ≤ α(s + u) − β(u) and the latter term is bounded by (α β)(s) by definition of the operator.


24

Now relax the assumption that the the inf in the definition of R ⊗ β is a min. In this case, the proof is essentially the same with a minor complication. For all > 0 there is some u ≥ 0 such that 0 ≤ t − s − u and (R ⊗ β)(t − s) ≥ R(t − s − u) + β(u) − and the proof continues along the same line, leading to: R∗ (t) − R∗ (t − s) ≤ (α β)(s) + This is true for all > 0, which proves the result.

a * a

d a ta x = b + rT b + rT

d = T + b /R 0

b

s lo p e R

b

e r s lo p

tim e

T

Figure 1.10: Computation of buffer, delay and output bounds for an input flow constrained by one leaky bucket, served in one node offered a rate-latency service curve. If r ≤ R, then the buffer bound is x = b + rT , the delay bound is d = T + Rb and the burstiness of the flow is increased by rT . If r > R, the bounds are infinite. A SIMPLE E XAMPLE AND I NTERPRETATION OF L EAKY B UCKET Consider a flow constrained by one leaky bucket, thus with an arrival curve of the form α = γr,b , served in a node with the service curve guarantee βR,T . The alert reader will enjoy applying the three bounds and finding the results shown in Figure 1.10. Consider in particular the case T = 0, thus a flow constrained by one leaky bucket served at a constant rate R. If R ≥ r then the buffer required to serve the flow is b, otherwise, it is infinite. This gives us a common interpretation of the leaky bucket parameters r and b: r is the minimum rate required to serve the flow, and b is the buffer required to serve the flow at any constant rate ≥ r. E XAMPLE : VBR FLOW WITH RATE - LATENCY SERVICE CURVE Consider a VBR flow, defined by TSPEC (M, p, r, b). This means that the flow has α(t) = min(M +pt, rt+b) as an arrival curve (Section 1.2). Assume that the flow is served in one node that guarantees a service curve equal to the rate-latency function β = βR,T . This example is the standard model used in Intserv. Let us apply Theorems 1.4.1 and 1.4.2. Assume that R ≥ r, that is, the reserved rate is as large as the sustainable rate of the flow. From the convexity of the region between α and β (Figure 1.4.1), we see that the vertical deviation v = sups≥0 [α(s) − β(s)] is reached for at an angular point of either α or β. Thus v = max[α(T ), α(θ) − β(θ)] with θ = b−M p−r . Similarly, the horizontal distance is reached an angular point. In the figure, it is either the distance marked as AA or BB . Thus, the bound on delay d is given by

α(θ) M d = max + T − θ, +T R R After some max-plus algebra, we can re-arrange these results as follows.

1.4. NETWORK CALCULUS BASICS

25

P ROPOSITION 1.4.1 (I NTSERV MODEL , BUFFER AND DELAY BOUNDS ). Consider a VBR flow, with TSPEC (M, p, r, b), served in a node that guarantees to the flow a service curve equal to the rate-latency function β = βR,T . The buffer required for the flow is bounded by

+ b−M v = b + rT + −T [(p − R)+ − p + r] p−r The maximum delay for the flow is bounded by d=

M+

b−M p−r (p

R

d a ta

A

B M 0

+T

a A

a (q )

− R)+

q

B

b

tim e

T

Figure 1.11: Computation of buffer and delay bound for one VBR flow served in one Intserv node. We can also apply Theorem 1.4.3 and find an arrival curve α∗ for the output flow. We have α∗ = α (λR ⊗ δT ) = (α λR ) δT from the properties of (Chapter 3). Note that (f δT )(t) = f (t + T ) for all f (shift to the left). The computation of α λR is explained in Theorem 3.1.14 on Page 126: it consists in inverting time, and smoothing. Here, we give however a direct derivation, which is possible since α is concave. Indeed, for a concave α, define t0 as t0 = inf{t ≥ 0 : α (t) ≤ R} where α is the left-derivative, and assume that t0 < +∞. A concave function always has a left-derivative, except maybe at the ends of the interval where it is defined. Then by studying the variations of the function u → α(t + u) − Ru we find that (α λR )(s) = α(s) if s ≥ t0 , and (α λR )(s) = α(t0 ) + (s − t0 )R if s < t0 . Putting the pieces all together we see that the output function α∗ is obtained from α by • replacing α on [0, t0 ] by the linear function with slope R that has the same value as α for t = t0 , keeping the same values as α on [t0 , +∞[, • and shifting by T to the left. Figure 1.12 illustrates the operation. Note that the two operations can be performed in any order since ⊗ is commutative. Check that the operation is equivalent to the construction in Theorem 3.1.14 on Page 126. If we apply this to a VBR connection, we obtain the following result. P ROPOSITION 1.4.2 (I NTSERV MODEL , OUTPUT BOUND ). With the same assumption as in Proposition 1.4.1, the output flow has an arrival curve α∗ given by: α∗ (t) = b + r(T + t) if b−M p−r ≤ T then + , b + r(T + t) (p − R) else α∗ (t) = min (t + T )(p ∧ R) + M + b−M p−r


26 bits

departure curve

slope = R

arrival curve

time -T

t0-T

Figure 1.12: Derivation of arrival curve for the output of a flow served in a node with rate-latency service curve βR,T . A N ATM E XAMPLE Consider the example illustrated in Figure 1.13. The aggregate flow has as an arrival curve equal to the stair function 10v25,4 . The figure illustrates that the required buffer is 18 ATM cells and the maximum delay is 10 time slots. We know from Corollary 1.2.1 that a GCRA constraint is c e lls b

g

x

r ,b

1 , 8

1 0 u

2 5 , 4

d d

x tim e s lo ts 0

1 0

2 0

3 0

4 0

5 0

Figure 1.13: Computation of bounds for buffer x and delay d for an ATM example. An ATM node serves 10 ATM connections, each constrained with GCRA(25, 4) (counted in time slots). The node offers to the aggregate flow a service curve βR,T with rate R = 1 cell per time slot and latency T = 8 time slots. The figure shows that approximating the stair function 10v25,4 by an affine function γr,b results into an overestimation of the bounds. equivalent to a leaky bucket. Thus, each of the 10 connections is constrained by an affine arrival curve γr,b 1 4 with r = 25 = 0.04 and b = 1 + 25 = 1.16. However, if we take as an arrival curve for the aggregate flow the resulting affine function 10γr,b , then we obtain a buffer bound of 11.6 and a delay bound of 19.6. The affine function overestimates the buffer and delay bounds. Remember that the equivalence between stair function and affine function is only for a flow where the packet size is equal to the value of the step, which is clearly not the case for an aggregate of several ATM connections. A direct application of Theorem 1.4.3 shows that an arrival curve for the output flow is given by α0∗ (t) = α(t + T ) = v25,12 (t). In Chapter 2, we give a slight improvement to the bounds if we know that the service curve is a strict service


27

curve.

1.4.2

A RE THE B OUNDS T IGHT ?

We now examine how good the three bounds are. For the backlog and delay bounds, the answer is simple: T HEOREM 1.4.4. Consider the backlog and delay bounds in Theorems 1.4.1 and 1.4.2. Assume that • α is a “good” function (that is, namely, is wide-sense increasing, sub-additive and α(0) = 0) • β is wide-sense increasing and β(0) = 0 Then the bounds are tight. More precisely, there is one causal system with input flow R(t) and output flow R∗ (t), such that the input is constrained by α, offering to the flow a service curve β, and which achieves both bounds. A causal system means that R(t) ≤ R∗ (t). The theorem means that the backlog bound in Theorem 1.4.1 is equal to supt≥0 [R(t) − R∗ (t)], and the delay bound in Theorem 1.4.1 is equal to supt≥0 d(t). In the above, d(t) is the virtual delay defined in Definition 1.1.1. P ROOF : We build one such system R, R∗ by defining R = α, R∗ = min(α, β). The system is causal because R∗ ≤ α = R. Now consider some arbitrary time t. If α(t) < β(t) then R∗ (t) = R(t) = R(t) + β(0) Otherwise,

R∗ (t) = β(t) = R(0) + β(t)

In all cases, for all t there is some s ≤ t such that R∗ (t) ≥ R(t − s) + β(s), which shows the service curve property. Of course, the bounds are as tight as the arrival and service curves are. We have seen that a source such that R(t) = α(t) is called greedy. Thus, the backlog and delay bounds are worst-case bounds that are achieved for greedy sources. In practice, the output bound is also a worst-case bound, even though the detailed result is somehow less elegant. T HEOREM 1.4.5. Assume that 1. 2. 3. 4.

α is a “good” function (that is, is wide-sense increasing, sub-additive and α(0) = 0) α is left-continuous β is wide-sense increasing and β(0) = 0 αα is not bounded from above.

Then the output bound in Theorem 1.4.3 is tight. More precisely, there is one causal system with input flow R(t) and output flow R∗ (t), such that the input is constrained by α, offering to the flow a service curve β, and α∗ (given by Theorem 1.4.3) is the minimum arrival curve for R∗ . We know in particular from Section 1.2 that the first three conditions are not restrictive. Let us first discuss the meaning of the last condition. By definition of max-plus deconvolution: (αα)(t) = inf {α(t + s) − α(s)} s≥0


28

One interpretation of αα is as follows. Consider a greedy source, with R(t) = α(t); then (αα)(t) is the minimum number of bits arriving over an interval of duration t. Given that the function is wide-sense increasing, the last condition means that limt→+∞ (αα)(t) = +∞. For example, for a VBR source with T-SPEC (p, M, r, b) (Figure 1.5), we have (αα)(t) = rt and the condition is satisfied. The alert reader will easily be convinced that the condition is also true if the arrival curve is a stair function. The proof of Theorem 1.4.5 is a little technical and is left at the end of this chapter. We might wonder whether the output bound α∗ is a “good” function. The answer is no, since α∗ (0) is the backlog bound and is positive in reasonable cases. However, α∗ is sub-additive (the proof is easy and left to the reader) thus the modified function δ0 ∧ α∗ defined as α∗ (t) for t > 0 and 0 otherwise is a “good” function. If α is left-continuous, δ0 ∧ α∗ is even a “very good” function since we know from the proof of Theorem 1.4.5 that it is left-continuous.

1.4.3

C ONCATENATION

So far we have considered elementary network parts. We now come to the main result used in the concatenation of network elements. T HEOREM 1.4.6 (C ONCATENATION OF N ODES ). Assume a flow traverses systems S1 and S2 in sequence. Assume that Si offers a service curve of βi , i = 1, 2 to the flow. Then the concatenation of the two systems offers a service curve of β1 ⊗ β2 to the flow. P ROOF : Call R1 the output of node 1, which is also the input to node 2. The service curve property at node 1 gives R1 ≥ R ⊗ β1 and at node 2

R∗ ≥ R1 ⊗ β2 ≥ (R ⊗ β1 ) ⊗ β2 = R ⊗ (β1 ⊗ β2 )

E XAMPLES : Consider two nodes offering each a rate-latency service curve βRi ,Ti , i = 1, 2, as is commonly assumed with Intserv. A simple computation gives βR1 ,T1 ⊗ βR1 ,T1 = βmin(R1 ,R2 ),T1 +T2 Thus concatenating Intserv nodes amounts to adding the latency components and taking the minimum of the rates. We are now also able to give another interpretation of the rate-latency service curve model. We know that βR,T = (δT ⊗ λR )(t); thus we can view a node offering a rate-latency service curve as the concatenation of a guaranteed delay node, with delay T and a constant bit rate or GPS node with rate R. PAY B URSTS O NLY O NCE The concatenation theorem allows us to understand a phenomenon known as “Pay Bursts Only Once”. Consider the concatenation of two nodes offering each a rate-latency service curve βRi ,Ti , i = 1, 2, as is commonly assumed with Intserv. Assume the fresh input is constrained by γr,b . Assume that r < R1 and r < R2 . We are interested in the delay bound, which we know is a worst case. Let us compare the results obtained as follows. 1. by applying the network service curve; 2. by iterative application of the individual bounds on every node


29

The delay bound D0 can be computed by applying Theorem 1.4.2: D0 = with R = mini (Ri ) and T0 =

i Ti

b + T0 R

as seen above.

Now apply the second method. A bound on the delay at node 1 is (Theorem 1.4.2): D1 =

b + T1 R1

The output of the first node is constrained by α∗ , given by : α∗ (t) = b + r × (t + T1 ) A bound on the delay at the second buffer is: D2 = And thus D1 + D2 =

b + rT1 + T2 R2 b b + rT1 + + T0 R1 R2

It is easy to see that D0 < D1 + D2 . In other words, the bounds obtained by considering the global service curve are better than the bounds obtained by considering every buffer in isolation. Let us continue the comparison more closely. The delay through one node has the form Rb1 + T1 (for the first node). The element Rb1 is interpreted as the part of the delay due to the burstiness of the input flow, whereas T1 is due to the delay component of the node. We see that D1 + D2 contains twice an element of the form Rbi , whereas D0 contains it only once. We sometimes say that “we pay bursts only once”. Another 1 difference between D0 and D1 + D2 is the element rT R2 : it is due to the increase of burstiness imposed by node 1. We see that this increase of burstiness does not result into an increase of the overall delay. A corollary of Theorem 1.4.6 is also that the end-to-end delay bound does not depend on the order in which nodes are concatenated.

1.4.4

I MPROVEMENT OF BACKLOG B OUNDS

We give two cases where we can slightly improve the backlog bounds. T HEOREM 1.4.7. Assume that a lossless node offers a strict service curve β to a flow with arrival curve α. Assume that α(u0 ) ≤ β(u0 ) for some u0 > 0. Then the duration of the busy period is ≤ u0 . Furthermore, for any time t, the backlog R(t) − R∗ (t) satisfies R(t) − R∗ (t) ≤

sup

u:0≤u0} as an arrival curve. The proof of the theorem is given later in this section. Before, we discuss the implications. Item 1 says that

C o m b in e d S y s t e m R (t )

R * (t )

b , g

P

(L )

R ( t )

B it - b y - b it s y s t e m Figure 1.18: The scenario and notation in Theorem 1.7.1. appending a packetizer to a node does not increase the packet delay at this node. However, as we see later, packetization does increase the end-to-end delay. Consider again the example in Section 1.7.1. A simple look at the figure shows that the backlog (or required buffer) is increased by the packetization, as indicated by item 2. Item 4 tells us that the final output R has σ (t) = σ(t) + lmax 1t>0 as an arrival curve, which is consistent with our observation in Section 1.7.1 that R is not σ-smooth, even though R∗ is. We will see in Section 1.7.4 that there is a stronger result, in relation with the concept of “packetized greedy shaper”. Item 3 is the most important practical result in this section. It shows that packetizing weakens the service curve guarantee by one maximum packet length. For example, if a system offers a rate-latency service curve with rate R, then appending a packetizer to the system has the effect of increasing the latency by lmax R . Consider also the example in Figure 1.16. The combination of the trunk and the packetizer can be modeled as a system offering • a minimum service curve βc, lmax c • a maximum service curve λc P ROOF OF T HEOREM 1.7.1 1. For some t such that Ti ≤ t < Ti+1 we have R(t) = L(i) and thus sup

d(t) = d(Ti )

t∈[Ti ,Ti+1 )

now Combining the two shows that

d(Ti ) = Ti − Ti sup d(t) = sup(Ti − Ti ) t

i


44

2. The proof is a direct consequence of Equation (1.21). 3. The result on maximum service curve γ follows immediately from Equation (1.21). Consider now the minimum service curve property. Fix some time t and define i0 by Ti0 ≤ t < Ti0 +1 . For 1 ≤ i ≤ i0 and for Ti−1 ≤ s < T1 we have R(s) = R(Ti−1 ) and β is wide-sense increasing, thus inf

Ti−1 ≤s 0. By definition of the limit to the right, we can find some s ∈ (Ti−1 , Ti ) such that β(t − s) ≤ βr (t − Ti ) + . Now R(s) = Rl (Ti ) thus R (t) ≥ R(s) + β(t − s) − ≥ (R ⊗ β )(t) − This is true for all > 0 thus R (t) ≥ (R ⊗ β )(t), which proves that the service curve property holds for case 1. The proof for case 2 is immediate. 4. The proof is a direct consequence of Equation (1.21). E XAMPLE : CONCATENATION OF GPS NODES Consider the concatenation of the theoretical GPS node, with guaranteed rate R (see Section 1.3.1 on Page 18) and an L-packetizer. Assume this system receives a flow of variable length packets. This models a theoretical node that would work as a GPS node but is constrained to deliver entire packets. This is not very realistic, and we will see in Chapter 2 more realistic implementations of GPS, but this example is sufficient to explain one important effect of packetizers. By applying Theorem 1.7.1, we find that this node offers a rate-latency service curve βR, lmax . Now conR catenate m such identical nodes, as illustrated in Figure 1.19. The end-to-end service curve is the rate latency-function βR,T with lmax T =m R We see on this example that the additional latency introduced by one packetizer is indeed of the order of one packet length; however, this effect is multiplied by the number of hops.

1.7. HANDLING VARIABLE LENGTH PACKETS

R a t e R , la t e n c y ( m

G P S

P

R

G P S L

45

1 ) lm

P

R

a x

/ R

G P S L

P

R

L

Figure 1.19: The concatenation of several GPS fluid nodes with packetized outputs For the computation of the end-to-end delay bound, we need to take into account Theorem 1.7.1, which tells us that we can forget the last packetizer. Thus, a bound on end-to-end delay is obtained by considering that the end-to-end path offers a service curve equal to the latency-function βR,T0 with T0 = (m − 1)

lmax R

For example, if the original input flow is constrained by one leaky bucket of rate r and bucket pool of size b, then an end-to-end delay bound is b + (m − 1)lmax (1.23) R The alert reader will easily show that this bound is a worst case bound. This illustrates that we should be careful in interpreting Theorem 1.7.1. It is only at the last hop that the packetizer implies no delay increase. The interpretation is as follows. Packetization delays the first bits in a packet, which delays the processing at downstream nodes. This effect is captured in Equation (1.23). In summary: R EMARK 1.7.1. Packetizers do not increase the maximum delay at the node where they are appended. However, they generally increase the end-to-end delay. We will see in Chapter 2 that many practical schedulers can be modeled as the concatenation of a node offering a service curve guarantee and a packetizer, and we will give a practical generalization of Equation (1.23).

1.7.3

A R ELATION BETWEEN G REEDY S HAPER AND PACKETIZER

We have seen previously that appending a packetizer to a greedy shaper weakens the arrival curve property of the output. There is however a case where this is not true. This case is important for the results in Section 1.7.4, but also has practical applications of its own. Figure 1.20 illustrates the theorem.

R

(P 0

(t )

L

)

(s ) R (t )

(P L

R * (t )

) R

Figure 1.20: Theorem 1.7.2 says that R(1) is σ-smooth.

(1 )

(t )


46

T HEOREM 1.7.2. Consider a sequence L of cumulative packet lengths and call PL the L-packetizer. Consider a “good” function σ and assume that There is a sub-additive function σ0 and a number l ≥ lmax such that (1.24) σ(t) = σ0 (t) + l1t>0 Call Cσ the greedy shaper with shaping curve σ. For any input, the output of the composition6 PL ◦ Cσ ◦ PL is σ-smooth. In practical terms, the theorem is used as follows. Consider an L-packetized flow, pass it through a greedy shaper with shaping curve σ; and packetize the output; then the result is σ-smooth (assuming that σ satisfies condition in Equation (1.24) in the theorem). Note that in general the output of Cσ ◦PL is not L-packetized, even if σ satisfies the condition in the theorem (finding a counter-example is simple and is left to the reader for her enjoyment). Similarly, if the input to PL ◦ Cσ is not L-packetized, then the output is not σ-smooth, in general. The theorem could also be rephrased by saying that, under condition in Equation (1.24) PL ◦ Cσ ◦ PL = Cσ ◦ PL ◦ Cσ ◦ PL since the two above operators always produce the same output. D ISCUSSION OF C ONDITION IN E QUATION (1.24) Condition Equation (1.24) is satisfied in practice if σ is concave and σr (0) ≥ lmax , where σr (0) = inf t>0 σ(t) is the limit to the right of σ at 0. This occurs for example if the shaping curve is defined by the conjunction of leaky buckets, all with bucket size at least as large as the maximum packet size. This also sheds some light on the example in Figure 1.16: the problem occurs because the shaping curve λC does not satisfy the condition. The alert reader will ask herself whether a sufficient condition for Equation (1.24) to hold is that σ is sub-additive and σr (0) ≥ lmax . Unfortunately, the answer is no. Consider for example the stair function σ = lmax vT . We have σr (0) = lmax but if we try to rewrite σ into σ(t) = σ0 (t) + l1t>0 we must have l = lmax and σ0 (t) = 0 for t ∈ (0, T ]; if we impose that σ0 is sub-additive, the latter implies σ0 = 0 which is not compatible with Equation (1.24).7 P ROOF OF T HEOREM 1.7.2: We use the notation in Figure 1.20. We want to show that R(1) is σsmooth. We have R∗ = R ⊗ σ. Consider now some arbitrary s and t with s < t. From the definition of min-plus convolution, for all > 0, there is some u ≤ s such that (R ⊗ σ)(s) ≥ R(u) + σ(s − u) −

(1.25)

Now consider the set E of > 0 such that we can find one u < s satisfying the above equation. Two cases are possible: either 0 is an accumulation point for E 8 (case 1) , or not (case 2). Consider case 1; there is a sequence (n , sn ), with sn < s, lim n = 0

n→+∞

and (R ⊗ σ)(s) ≥ R(sn ) + σ(s − sn ) − n We use the notation PL ◦ Cσ to denote the composition of the two operators, with Cσ applied first; see Section 4.1.3. The same conclusion unfortunately also holds if we replace sub-additive by “star-shaped” (Section 3.1). 8 namely, there is a sequence of elements in E which converges to 0 6 7


47

Now since sn ≤ t: (R ⊗ σ)(t) ≤ R(sn ) + σ(t − sn ) Combining the two: (R ⊗ σ)(t) − (R ⊗ σ)(s) ≤ σ(t − sn ) − σ(s − sn ) + n Now t − sn > 0 and s − sn > 0 thus σ(t − sn ) − σ(s − sn ) = σ0 (t − sn ) − σ0 (s − sn ) We have assumed that σ0 is sub-additive. Now t ≥ s thus σ0 (t − sn ) − σ0 (s − sn ) ≤ σ0 (t − s) we have thus shown that, for all n (R ⊗ σ)(t) − (R ⊗ σ)(s) ≤ σ0 (t − s) + n and thus (R ⊗ σ)(t) − (R ⊗ σ)(s) ≤ σ0 (t − s) Now from Equation (1.21), it follows that R(1) (t) − R(1) (s) ≤ σ0 (t − s) + lmax ≤ σ(t − s) which ends the proof for case 1. Now consider case 2. There is some 0 such that for 0 < < 0 , we have to take u = s in Equation (1.25). This implies that (R ⊗ σ)(s) = R(s) Now R is L-packetized by hypothesis. Thus R(1) (s) = P L ((R ⊗ σ)(s)) = P L (R(s)) = R(s) = (R ⊗ σ)(s) thus

R(1) (t) − R(1) (s) = P L ((R ⊗ σ)(t) − (R ⊗ σ)(s) ≤ (R ⊗ σ)(t) − (R ⊗ σ)(s)

now R ⊗ σ has σ as an arrival curve thus R(1) (t) − R(1) (s) ≤ σ(t − s) which ends the proof for case 2. E XAMPLE : B UFFERED L EAKY B UCKET C ONTROLLER BASED ON V IRTUAL F INISH T IMES Theorem 1.7.2 gives us a practical implementation for a packet based shaper. Consider that we want to build a device that ensures that a packet flow satisfies some concave, piecewise linear arrival curve (and is of course L- packetized). We can realize such a device as the concatenation of a buffered leaky bucket controller operating bit-by-bit and a packetizer. We compute the output time for the last bit of a packet (= finish time) under the bit-by-bit leaky bucket controller, and release the entire packet instantly at this finish time. If each bucket pool is at least as large as the maximum packet size then Theorem 1.7.2 tells us that the final output satisfies the leaky bucket constraints.


48 1 0 0

R *

7 5

5 0

2 5

R 1

2

3

(1 ) 4

5

t

Figure 1.21: A counter example for Theorem 1.7.2. A burst of 10 packets of size equal to 10 data units arrive at time t = 0, and σ = 25v1 . The greedy shaper emits 25 data units at times 0 and 1, which forces the packetizer to create a burst of 3 packets at time 1, and thus R(1) is not σ-smooth.

C OUNTER - EXAMPLE If we consider non-concave arrival curves, then we can find an arrival curve σ that does satisfy σ(t) ≥ lmax for t > 0 but that does not satisfy Equation (1.24). In such a case, the conclusion of Theorem 1.7.2 may not hold in general. Figure 1.21 shows an example where the output R(1) is not σ-smooth, when σ is a stair function.

1.7.4

PACKETIZED G REEDY S HAPER

We can come back to the questions raised by the example in Figure 1.16 and give a more fundamental look at the issue of packetized shaping. Instead of synthesizing the concatenation of a greedy shaper and a packetizer as we did earlier, we define the following, consistent with Section 1.5. D EFINITION 1.7.6. [Packetized Greedy Shaper] Consider an input sequence of packets, represented by the function R(t) as in Equation (1.18). Call L the cumulative packet lengths. We call packetized shaper, with shaping curve σ, a system that forces its output to have σ as an arrival curve and be L-packetized. We call packetized greedy shaper a packetized shaper that delays the input packets in a buffer, whenever sending a packet would violate the constraint σ, but outputs them as soon as possible.

E XAMPLE : B UFFERED L EAKY B UCKET C ONTROLLER BASED ON B UCKET R EPLENISHMENT The case σ(t) = minm=1,...,M (γrm ,bm (t) can be implemented by a controller that observes a set of M fluid buckets, where the mth bucket is of size bm and leaks at a constant rate rm . Every bucket receives li units of fluid when packet i is released (li is the size of packet i). A packet is released as soon as the level of fluid in bucket m allows it, that is, has gone down below bm − li , for all m. We say that now we have defined a buffered leaky bucket controller based on “bucket replenishment”. It is clear that the output has σ as an arrival curve, is L-packetized and sends the packets as early as possible. Thus it implements the packetized greedy shaper. Note that this implementation differs from the buffered leaky bucket controller based on virtual finish times introduced in Section 1.7.3. In the latter, during a period where, say, bucket m only is full, fragments of a packet are virtually released at rate rm , bucket m remains full, and the (virtual) fragments are


49

then re-assembled in the packetizer; in the former, if a bucket becomes full, the controller waits until it empties by at least the size of the current packet. Thus we expect that the level of fluid in both systems is not the same, the former being an upper bound. We will see however in Corollary 1.7.1 that both implementations are equivalent. In this example, if a bucket size is less than the maximum packet size, then it is never possible to output a packet: all packets remain stuck in the packet buffer, and the output is R(t) = 0. In general, we can say that P ROPOSITION 1.7.1. If σr (0) < lmax then the the packetized greedy shaper blocks all packets for ever (namely, R(t) = 0). Thus in this section, we assume that σ(t) ≥ lmax for t > 0. Thus, for practical cases, we have to assume that the arrival curve σ has a discontinuity at the origin at least as large as one maximum packet size. How does the packetized greedy shaper compare with the concatenation of a greedy shaper with shaping curve σ and a packetizer ? We know from the example in Figure 1.16 that the output has σ (t) = σ(t) + lmax 1t>0 as an arrival curve, but not σ. Now, does the concatenation implement a packetized greedy shaper with shaping curve σ ? Before giving a general answer, we study a fairly general consequence of Theorem 1.7.2. T HEOREM 1.7.3 (R EALIZATION OF PACKETIZED G REEDY S HAPER ). Consider a sequence L of cumulative packet lengths and a “good” function σ. Assume that σ satisfies the condition in Equation (1.24). Consider only inputs that are L packetized. Then the packetized greedy shaper for σ and L can be realized as the concatenation of the greedy shaper with shaping curve σ and the L-packetizer.

P a c k e t iz e d G r e e d y S h a p e r (L ) a n d s (s )

(P L

)

Figure 1.22: The packetized greedy shaper can be realized as a (bit-by-bit fluid shaper followed by a packetizer, assuming Equation (1.24) holds. In practice, this means that we can realize packetized greedy shaping by computing finish times in the virtual fluid system and release packets at their finish times.

P ROOF : Call R(t) the packetized input; the output of the bit-by-bit greedy shaper followed by a packetizer is R(1) (t) = P L (R ⊗ σ)(t)). Call R(t) the output of the packetized greedy shaper. We have R ≤ R thus R ⊗ σ ≤ R ⊗ σ and thus P L (R ⊗ σ) ≤ P L (R ⊗ σ) But R is σ-smooth, thus R ⊗ σ = R, and is L-packetized, thus P L (R ⊗ σ) = R. Thus the former inequality can be rewritten as R ≤ R(1) . Conversely, from Theorem 1.7.2, R(1) is also σ-smooth and L-packetized. The definition of the packetized greedy shaper implies that R ≥ R(1) (for a formal proof, see Lemma 1.7.1) thus finally R = R(1) .


50

We have seen that the condition in the theorem is satisfied in particular if σ is concave and σr (0) ≥ lmax , for example if the shaping curve is defined by the conjunction of leaky buckets, all with bucket size at least as large as the maximum packet size. This shows the following. C OROLLARY 1.7.1. For L-packetized inputs, the implementations of buffered leaky bucket controllers based on bucket replenishment and virtual finish times are equivalent. If we relax Equation (1.24) then the construction of the packetized greedy shaper is more complex: T HEOREM 1.7.4 (I/O CHARACTERISATION OF PACKETIZED GREEDY SHAPERS ). Consider a packetized greedy shaper with shaping curve σ and cumulative packet length L. Assume that σ is a “good” function. The output R(t) of the packetized greedy shaper is given by

R = inf R(1) , R(2) , R(3) , ...

(1.26)

with R(1) (t) = P L ((σ ⊗ R)(t)) and R(i) (t) = P L ((σ ⊗ R(i−1) )(t)) for i ≥ 2. Figure 1.23 illustrates the theorem, and shows the iterative construction of the output on one example. Note that this example is for a shaping function that does not satisfy Equation (1.24). Indeed, otherwise, we know from Theorem 1.7.3 that the iteration stops at the first step, namely, R = R(1) in that case. We can also check for example that if σ = λr (thus the condition in Proposition 1.7.1 is satisfied) then the result of Equation (1.26) is 0. 0

R R

in f P L

1

2

3

4

5

(1 )

R

(2 )

R

(3 )

R = R

(4 )

s

Figure 1.23: Representation of the output of the packetized greedy shaper (left) and example of output (right). The data are the same as with Figure 1.21.

P ROOF : The proof is a direct application of Lemma 1.7.1 (which itself is an application of the general method in Section 4.3 on Page 144).


51

L EMMA 1.7.1. Consider a sequence L of cumulative packet lengths and a “good” function σ. Among all flows x(t) such that ⎧ ⎨ x≤R x is L-packetized (1.27) ⎩ x has σ as an arrival curve there is one flow R(t) that upper-bounds all. It is given by Equation (1.26). P ROOF : The lemma is a direct application of Theorem 4.3.1, as explained in Section 4.3.2. However, in order to make this chapter self-contained, we give an alternative, direct proof, which is quite short. If x is a solution, then it is straightforward to show by induction on i that x(t) ≤ R(i) (t) and thus x ≤ R. The difficult part is now to show that R is indeed a solution. We need to show that the three conditions in Equation (1.27) hold. Firstly, R(1) ≤ R(t) and by induction on i, R(i) ≤ R for all i; thus R ≤ R. Secondly, consider some fixed t; R(i) (t) is L-packetized for all i ≥ 1. Let L(n0 ) := R(1) (t). Since R(i) (t) ≤ R(1) (t), R(i) (t) is in the set {L(0), L(1), L(2), ..., L(n0 )}. This set is finite, thus, R(t), which is the infimum of elements in this set, has to be one of the L(k) for k ≤ n0 . This shows that R(t) is L-packetized, and this is true for any time t. Thirdly, we have, for all i R(t) ≤ R(i+1) (t) = P L ((σ ⊗ R(i) )(t)) ≤ (σ ⊗ R(i) )(t) thus R ≤ inf (σ ⊗ R(i) ) i

Now convolution by a fixed function is upper-semi-continuous, which means that inf (σ ⊗ R(i) ) = σ ⊗ R i

This is a general result in Chapter 4 for any min-plus operator. An elementary proof is as follows. (i) inf i (σ ⊗ R(i) )(t) = inf s∈[0,t],i∈N σ(s) + R (t −(i)s) R (t − s) = inf s∈[0,t] inf i∈N (σ(s) + = inf s∈[0,t] σ(s) + inf i∈N R (i) (t − s) = inf s∈[0,t] σ(s) + R(t − s) = (σ ⊗ R)(t) Thus R ≤ σ ⊗ R, which shows the third condition. Note that R is wide-sense increasing. D OES A PACKETIZED GREEDY SHAPER KEEP ARRIVAL CONSTRAINTS ? Figure 1.24 shows a counterexample, namely, a variable length packet flow that has lost its initial arrival curve constraint after traversing a packetized greedy shaper. However, if arrival curves are defined by leaky buckets, we have a positive result. T HEOREM 1.7.5 (C ONSERVATION OF CONCAVE ARRIVAL CONSTRAINTS ). Assume an L-packetized flow with arrival curve α is input to a packetized greedy shaper with cumulative packet length L and shaping curve σ. Assume that α and σ are concave with αr (0) ≥ lmax and σr (0) ≥ lmax . Then the output flow is still constrained by the original arrival curve α.


52

0

1

2

3

4

5

R R Figure 1.24: The input flow is shown above; it consists of 3 packets of size 10 data units and one of size 5 data units, spaced by one time unit. It is α-smooth with α = 10v1,0 . The bottom flow is the output of the packetized greedy shaper with σ = 25v3,0 . The output has a burst of 15 data units packets at time 3. It is σ-smooth but not α-smooth. P ROOF : Since σ satisfies Equation (1.24), it follows from Theorem 1.7.3 that R = P L (σ ⊗ R). Now R is α-smooth thus it is not modified by a bit-by-bit greedy shaper with shaping curve α, thus R = α ⊗ R. Combining the two and using the associativity of ⊗ gives R = P L [(σ ⊗ α) ⊗ R]. From our hypothesis, σ ⊗ α = min(σ, α) (see Theorem 3.1.6 on Page 112) and thus σ ⊗ α satisfies Equation (1.24). Thus, by Theorem 1.7.2, R is σ ⊗ α-smooth, and thus α-smooth. S ERIES DECOMPOSITION OF SHAPERS T HEOREM 1.7.6. Consider a tandem of M packetized greedy shapers in series; assume that the shaping curve σ m of the mth shaper is concave with σrm (0) ≥ lmax . For L-packetized inputs, the tandem is equivalent to the packetized greedy shaper with shaping curve σ = minm σ m . P ROOF : We do the proof for M = 2 as it extends without difficulty to larger values of M . Call R(t) the packetized input, R (t) the output of the tandem of shapers, and R(t) the output of the packetized greedy shaper with input R(t). Firstly, by Theorem 1.7.3 Now σ m ≥ σ for all m thus

R = P L [σ 2 ⊗ P L (σ 1 ⊗ R)] R ≥ P L [σ ⊗ P L (σ ⊗ R)]

Again by Theorem 1.7.3, we have R = P L (σ ⊗ R). Moreover R is L-packetized and σ-smooth, thus R = P L (R) and R = σ ⊗ R. Thus finally (1.28) R ≥ R Secondly, R is L-packetized and by Theorem 1.7.5, it is σ-smooth. Thus the tandem is a packetized (possibly non greedy) shaper. Since R(t) is the output of the packetized greedy shaper, we must have R ≤ R. Combining with Equation (1.28) ends the proof. It follows that a shaper with shaping curve σ(t) = minm=1,...,M (rm t + bm ), where bm ≥ lmax for all m, can be implemented by a tandem of M individual leaky buckets, in any order. Furthermore, by Corollary 1.7.1, every individual leaky bucket may independently be based either on virtual finish times or on bucket replenishment. If the condition in the theorem is not satisfied, then the conclusion may not hold. Indeed, for the example in Figure 1.24, the tandem of packetized greedy shapers with curves α and σ does not have an α-smooth output, therefore it cannot be equivalent to the packetized greedy shaper with curve min(α, σ).

1.8. EFFECTIVE BANDWIDTH AND EQUIVALENT CAPACITY

53

Unfortunately, the other shaper properties seen in Section 1.5 do not generally hold. For shaping curves that satisfy Equation (1.24), and when a packetized greedy shaper is introduced, we need to compute the end-to-end service curve by applying Theorem 1.7.1.

1.8 1.8.1

L OSSLESS E FFECTIVE BANDWIDTH AND E QUIVALENT C APACITY E FFECTIVE BANDWIDTH OF A F LOW

We can apply the results in this chapter to define a function of a flow called the effective bandwidth. This function characterizes the bit rate required for a given flow. More precisely, consider a flow with cumulative function R; for a fixed, but arbitrary delay D, we define the effective bandwidth eD (R) of the flow as the bit rate required to serve the flow in a work conserving manner, with a virtual delay ≤ D. P ROPOSITION 1.8.1. The effective bandwidth of a flow is given by R(t) − R(s) 0≤s≤t t − s + D

eD (R) = sup

(1.29)

For an arrival curve α we define the effective bandwidth eD (α) as the effective bandwidth of the greedy flow R = α. By a simple manipulation of Equation 1.29, the following comes. P ROPOSITION 1.8.2. The effective bandwidth of a “good” arrival curve is given by eD (α) = sup 0≤s

α(s) s+D

(1.30)

The alert reader will check that the effective bandwidth of a flow R is also the effective bandwidth of its minimum arrival curve R R. For example, for a flow with T-SPEC (p, M, r, b), the effective bandwidth is the maximum of r and the slopes of lines (QA0 ) and (QA1 ) in Figure 1.25; it is thus equal to: D− M M p (1.31) , r, p 1 − b−M eD = max D + D p−r Assume α is sub-additive. We define the sustainable rate m as m = lim inf s→+∞

α(s) s

and the peak rate by

a r r iv a l c u r v e A

b M Q

s lo p e r 1

A

s lo p e p 0

2 0 0 1 0 0 5 0 2 0

0 .0 5

0 .1

0 .2

0 .5

1

2

Figure 1.25: Computation of Effective Bandwidth for a VBR flow (left); example for r = 20 packets/second, M = 10 packets, p = 200 packets per second and b = 26 packets (right). p = sups>0 α(s) s . Then m ≤ eD (α) ≤ p for all D. Moreover, if α is concave, then limD→+∞ eD (α) = m.If α is differentiable, e(D) is the slope of the tangent to the arrival curve, drawn from the time axis at t = −D (Figure 1.26). It follows also directly from the definition in (1.29) that


54 bits

slope = effective bandwidth

bits

arrival curve

slope = equivalent capacity arrival curve

B time -D

Figure 1.26: Effective Bandwidth for a delay constraint D and Equivalent Capacity for a buffer size B eD ( αi ) ≤ eD (αi ) i

(1.32)

i

In other words, the effective bandwidth for an aggregate flow is less than or equal to the sum of effective bandwidths. If the flows have all identical arrival curves, then the aggregate effective bandwidth is simply this latter relation that is the origin of the term “effective bandwidth”. The difference I× eD (α1 ). It is e (α ) − e ( i D D i i αi ) is a buffering gain; it tells us how much capacity is saved by sharing a buffer between the flows.

1.8.2

E QUIVALENT C APACITY

Similar results hold if we replace delay constraints by the requirement that a fixed buffer size is not exceeded. Indeed, the queue with constant rate C, guarantees a maximum backlog of B (in bits) for a flow R if C ≥ fB (R), with R(t) − R(s) − B (1.33) fB (R) = sup t−s 0≤s0

α(s) − B s

(1.34)

We call fB (α) the equivalent capacity, by analogy to [48]. Similar to effective bandwidth, the equivalent capacity of a heterogeneous mix of flows is less than or equal to the sum ofequivalent capacities of the flows, provided that the buffers are also added up; in other words, f (α) ≤ f (α ), with α = i B i Bi i αi and B = i Bi . Figure 1.26 gives a graphical interpretation. For example, for a flow with T-SPEC (p, M, r, b), using the same method as above, we find the following equivalent capacity: if B < M then + ∞ + (1.35) fB = else r + (p−r)(b−B) b−M An immediate computation shows that fb (γr,b ) = r. In other words, if we allocate to a flow, constrained by an affine function γr,b , a capacity equal to its sustainable rate r, then a buffer equal to its burst tolerance b is sufficient to ensure loss-free operation. Consider now a mixture of Intserv flows (or VBR connections), with T-SPECs (Mi , pi , ri , bi ). If we allocate to this aggregate of flows the sum of their sustainable rates i ri , then the buffer requirement is the sum of

1.8. EFFECTIVE BANDWIDTH AND EQUIVALENT CAPACITY

55

the burst tolerances i bi , regardless of other parameters such as peak rate. Conversely, Equation 1.35 also illustrates that there is no point allocating more buffer than the burst tolerance: if B > b, then the equivalent capacity is still r. The above has illustrated that it is possible to reduce the required buffer or delay by allocating a rate larger than the sustainable rate. In Section 2.2, we described how this may be done with a protocol such as RSVP. Note that formulas (1.29) or (1.33), or both, can be used to estimate the capacity required for a flow, based on a measured arrival curve. We can view them as low-pass filters on the flow function R.

1.8.3

E XAMPLE : ACCEPTANCE R EGION FOR A FIFO M ULTIPLEXER

Consider a node multiplexing n1 flows of type 1 and n2 flows of type 2, where every flow is defined by a T-SPEC (pi , Mi , ri , bi ). The node has a constant output rate C. We wonder how many flows the node can accept. If the only condition for flow acceptance is that the delay for all flows is bounded by some value D, then the set of acceptable values of (n1 , n2 ) is defined by eD (n1 α1 + n2 α2 ) ≤ C We can use the same convexity arguments as for the derivation of formula (1.31), applied to the function n1 α1 + n2 α2 . Define θi = bpii−M −ri and assume θ1 ≤ θ2 . The result is: ⎧ ⎪ ⎪ ⎪ ⎨ eD (n1 α1 + n2 α2 ) = max

⎪ ⎪ ⎪ ⎩

n1 M1 +n2 M2 , D n1 M1 +n2 M2 +(n1 p1 +n2 p2 )θ1 , θ1 +D n1 b1 +n2 M2 +(n1 r1 +n2 p2 )θ2 , θ2 +D

n 1 r1 + n 2 r2

The set of feasible (n1 , n2 ) derives directly from the previous equation; it is the convex part shown in Figure 1.27. The alert reader will enjoy performing the computation of the equivalent capacity for the case where the acceptance condition bears on a buffer size B. i 1 2

pi 20’000 packets/s 5’000 packets/s

Mi 1 packet 1 packet

ri 500 packets/s 500 packets/s

bi 26 packets 251 packets

θi 1.3 ms 55.5 ms

Figure 1.27: Acceptance region for a mix of type 1 and type 2 flows. Maximum delay D = xx. The parameters for types 1 and 2 are shown in the table, together with the resulting values of θi . Coming back to equation 1.32, we can state in more general terms that the effective bandwidth is a convex function of function α, namely: eD (aα1 + (1 − a)α2 ) ≤ aeD (α1 ) + (1 − a)eD (α2 ) for all a ∈ [0, 1]. The same is true for the equivalent capacity function. Consider now a call acceptance criterion based solely on a delay bound, or based on a maximum buffer constraint, or both. Consider further that there are I types of connections, and define the acceptance region A as the set of values (n1 , . . . , nI ) that satisfy the call acceptance criterion, where ni is the number of connections of class i. From the convexity of the effective bandwidth and equivalent capacity functions, it follows that the acceptance region A is convex. In chapter 9 we compare this to acceptance regions for systems with some positive loss probability.


56

S USTAINABLE R ATE A LLOCATION If we are interested only in course results, then we can reconsider the previous solution and take into account only the sustainable rate mix. The aggregate of the connection flow is constrained (among others) by α(s) = b + rs, with b = i ni bi and r = i ni ri . Theorem 1.4.1 shows that the maximum aggregate buffer occupancy is bounded by b as long as C ≥ r. In other words, allocating the sustainable rate guarantees a loss-free operation, as long as the total buffer is equal to the burstiness. In a more general setting, assume an aggregate flow has α as minimum arrival curve, and assume that some parameters r and b are such that lim α(s) − rs − b = 0 s→+∞

so that the sustainable rate r with burstiness b is a tight bound. It can easily be shown that if we allocate a rate C = r, then the maximum buffer occupancy is b. Consider now multiplexing a number of VBR connections. If no buffer is available, then it is necessary for a loss-free operation to allocate the sum of the peak rates. In contrast, using a buffer of size b makes it possible to allocate only the sustainable rate. This is what we call the buffering gain, namely, the gain on the peak rate obtained by adding some buffer. The buffering gain comes at the expense of increased delay, as can easily be seen from Theorem 1.4.2.

1.9

P ROOF OF T HEOREM 1.4.5

S TEP 1: Consider a fixed time t0 and assume, in this step, that there is some time u0 that achieves the supremum in the definition of α β. We construct some input and output functions R and R∗ such that R is constrained by α, the system (R, R∗ ) is causal, and α∗ (t0 ) = (R∗ R∗ )(t0 ). R and R∗ are given by (Figure 1.28) d a ta

a R b R *

0

u 0

u 0

+ t0

tim e

Figure 1.28: Step 1 of the proof of Theorem 1.4.5: a system that attains the output bound at one value t0 . ⎧ R(t) = α(t) if t < u0 + t0 ⎪ ⎪ ⎨ R(t) = α(u0 + t0 ) if t ≥ u0 + t0 R∗ (t) = inf[α(t), β(t)] if t < u0 + t0 ⎪ ⎪ ⎩ ∗ R (t) = R(t) if t ≥ u0 + t0 It is easy to see, as in the proof of Theorem 1.4.4 that R and R∗ are wide-sense increasing, that R∗ ≤ R and that β is a service curve for the flow. Now R∗ (u0 + t0 ) − R∗ (u0 ) = α(u0 + t0 ) − R∗ (u0 ) ≥ α(u0 + t0 ) − β(u0 ) = α∗ (t0 ) S TEP 2: Consider now a sequence of times t0 , t1 , ..., tn , ... (not necessarily increasing). Assume, in this step, that for all n there is a value un that achieves the supremum in the definition of (α β)(tn ). We prove that there are some functions R and R∗ such that R is constrained by α, the system (R, R∗ ) is causal, has β as a service curve, and α∗ (tn ) = (R∗ R∗ )(tn ) for all n ≥ 0.

1.9. PROOF OF THEOREM 1.4.5

57

We build R and R∗ by induction on a set of increasing intervals [0, s0 ], [0, s1 ],..., [0, sn ].... The induction property is that the system restricted to time interval [0, sn ] is causal, has α as an arrival curve for the input, has β as a service curve, and satisfies α∗ (ti ) = (R∗ R∗ )(ti ) for i ≤ n. The first interval is defined by s0 = u0 + t0 ; R and R∗ are built on [0, s0 ] as in step 1 above. Clearly, the induction property is true for n = 0. Assume we have built the system on interval [0, sn ]. Define now sn+1 = sn + un + tn + δn+1 . We chose δn+1 such that α(s + δn+1 ) − α(s) ≥ R(sn ) for all s ≥ 0

(1.36)

This is possible from the last condition in the Theorem. The system is defined on ]sn , sn+1 ] by (Figure 1.29) ⎧ R(t) = R∗ (t) = R(sn ) for sn < t ≤ sn + δn+1 ⎪ ⎪ ⎨ R(t) = R(sn ) + α(t − sn − δn+1 ) for sn + δn+1 < t ≤ sn+1 ⎪ R∗ (t) = R(sn ) + (α ∧ β)(t − sn − δn+1 ) for sn + δn+1 < t < sn+1 ⎪ ⎩ ∗ R (sn+1 ) = R(sn+1 ) We show now that the arrival curve constraint is satisfied for the system defined on [0, sn+1 ]. Consider a b R

d a ta

R * a b

a b d 0 u 0

s0= u 0

+ t0

1

u 1

d

t1 s

2

u 2

1

t2 s 2

tim e

Figure 1.29: Step 2 of the proof of Theorem 1.4.5: a system that attains the output bound for all values tn , n ∈ N. R(t) − R(v) for t and v in [0, sn+1 ]. If both t ≤ sn and v ≤ sn , or if both t > sn and v > sn then the arrival curve property holds from our construction and the induction property. We can thus assume that t > sn and v ≤ sn . Clearly, we can even assume that t ≥ sn + δn+1 , otherwise the property is trivially true. Let us rewrite t = sn + δn+1 + s. We have, from our construction: R(t) − R(v) = R(sn + δn+1 + s) − R(v) = R(sn ) + α(s) − R(v) ≤ R(sn ) + α(s) Now from Equation (1.36), we have: R(sn ) + α(s) ≤ α(s + δn+1 ) ≤ α(s + δn+1 + sn − v) = α(t − v) which shows the arrival curve property. Using the same arguments as in step 1, it is simple to show that the system is causal, has β as a service curve, and that R∗ (un+1 + tn+1 ) − R∗ (un+1 ) = α∗ (tn+1 ) which ends the proof that the induction property is also true for n + 1.


58

S TEP 3: Consider, as in step 2, a sequence of times t0 , t1 , ..., tn , ... (not necessarily increasing). We now extend the result in step 2 to the case where the supremum in the definition of α∗ = (α β)(tn ) is not necessarily attained. Assume first that α∗ (tn ) is finite for all n. For all n and all m ∈ N∗ there is some um,n such that 1 (1.37) α(tn + um,n ) − β(um,n ) ≥ α∗ (tn ) − m Now the set of all couples (m, n) is enumerable. Consider some numbering (M (i), N (i)), i ∈ N for that set. Using the same construction as in step 2, we can build by induction on i a sequence of increasing intervals [0, si ] and a system (R, R∗ ) that is causal, has α as an arrival curve for the input, has β as a service curve, and such that 1 R∗ (si ) − R∗ (si − tN (i) ) ≥ α∗ (tN (i) ) − M (i) Now consider an arbitrary, but fixed n. By applying the previous equations to all i such that N (i) = n, we obtain

(R∗ R∗ )(tn ) ≥ supi such that N (i)=n α∗ (tN (i) ) − M1(i)

= α∗ (tn ) − inf i such that N (i)=n M1(i) Now the set of all

1 M (i)

for i such that N (i) = n is N∗ , thus 1 =0 inf i such that N (i)=n M (i)

and thus (R∗ R∗ )(tn ) = α∗ (tn ), which ends the proof of step 3 in the case where α∗ (tn ) is finite for all n. A similar reasoning can be used if α∗ (tn ) is infinite for some tn . In that case replace Equation (1.37) by α(tn + um,n ) − β(um,n ) ≥ m. S TEP 4: Now we conclude the proof. If time is discrete, then step 3 proves the theorem. Otherwise we use a density argument. The set of nonnegative rational numbers Q+ is enumerable; we can thus apply step 3 to the sequence of all elements of Q+ , and obtain system (R, R∗ ), with (R∗ R∗ )(q) = α∗ (q) for all q ∈ Q+ Function R∗ is right-continuous, thus, from the discussion at the end of Theorem 1.2.2, it follows that R∗ R∗ is left-continuous. We now show that α∗ is also left-continuous. For all t ≥ 0 we have: sup α∗ (s) = s r for all t ? E XERCISE 1.17. Consider a series of guaranteed service nodes with service curves ci (t) = ri (t − Ti )+ . What is the maximum delay through this system for a flow constrained by (m, b) ?

1.11. EXERCISES

63

E XERCISE 1.18. A flow with T-SPEC (p, M, r, b) traverses nodes 1 and 2. Node i offers a service curve ci (t) = Ri (t − Ti )+ . What buffer size is required for the flow at node 2 ? E XERCISE 1.19. A flow with T-SPEC (p, M, r, b) traverses nodes 1 and 2. Node i offers a service curve ci (t) = Ri (t − Ti )+ . A shaper is placed between nodes 1 and 2. The shaper forces the flow to the arrival curve z(t) = min(R2 t, bt + m). 1. What buffer size is required for the flow at the shaper ? 2. What buffer size is required at node 2 ? What value do you find if T1 = T2 ? 3. Compare the sum of the preceding buffer sizes to the size that would be required if no re-shaping is performed. 4. Give an arrival curve for the output of node 2. E XERCISE 1.20. Prove the formula giving of paragraph “Buffer Sizing at a Re-shaper” E XERCISE 1.21. Is Theorem “Input-Output Characterization of Greedy Shapers” a stronger result than Corollary “Service Curve offered by a Greedy Shaper” ? E XERCISE 1.22. 1. Explain what is meant by “we pay bursts only once”. 2. Give a summary in at most 15 lines of the main properties of shapers 3. Define the following concepts by using the ⊗ operator: Service Curve, Arrival Curve, Shaper 4. What is a greedy source ? E XERCISE 1.23.

1. Show that for a constant bit rate trunk with rate c, the backlog at time t is given by W (t) = sup {R(t) − R∗ (s) − c(t − s)} s≤t

2. What does the formula become if we assume only that, instead a constant bit rate trunk, the node is a scheduler offering β as a service curve ? E XERCISE 1.24. Is it true that offering a service curve β implies that, during any busy period of length t, the amount of service received rate is at least β(t) ? E XERCISE 1.25. A flow S(t) is constrained by an arrival curve α. The flow is fed into a shaper, with shaping curve σ. We assume that α(s) = min(m + ps, b + rs) and σ(s) = min(P s, B + Rs) We assume that p > r, m ≤ b and P ≥ R. The shaper has a fixed buffer size equal to X ≥ m. We require that the buffer never overflows. 1. Assume that B = +∞. Find the smallest of P which guarantees that there is no buffer overflow. Let P0 be this value. 2. We do not assume that B = +∞ any more, but we assume that P is set to the value P0 computed in the previous question. Find the value (B0 , R0 ) of (B, R) which guarantees that there is no buffer overflow and minimizes the cost function c(B, R) = aB + R, where a is a positive constant. What is the maximum virtual delay if (P, B, R) = (P0 , B0 , R0 ) ? E XERCISE 1.26. We consider a buffer of size X cells, served at a constant rate of c cells per second. We put N identical connections into the buffer; each of the N connections is constrained both by GCRA(T1 , τ1 ) and GCRA(T2 , τ2 ). What is the maximum value of N which is possible if we want to guarantee that there is no cell loss at all ? Give the numerical application for T1 = 0.5 ms, τ1 = 4.5 ms, T2 = 5 ms, τ2 = 495 ms, c = 106 cells/second, X = 104 cells


64

E XERCISE 1.27. We consider a flow defined by its function R(t), with R(t) = the number of bits observed since time t = 0. 1. The flow is fed into a buffer, served at a rate r. Call q(t) the buffer content at time t. We do the same assumptions as in the lecture, namely, the buffer is large enough, and is initially empty. What is the expression of q(t) assuming we know R(t) ? We assume now that, unlike what we saw in the lecture, the initial buffer content (at time t = 0) is not 0, but some value q0 ≥ 0. What is now the expression for q(t) ? 2. The flow is put into a leaky bucket policer, with rate r and bucket size b. This is a policer, not a shaper, so nonconformant bits are discarded. We assume that the bucket is large enough, and is initially empty. What is the condition on R which ensures that no bit is discarded by the policer (in other words, that the flow is conformant) ? We assume now that, unlike what we saw in the lecture, the initial bucket content (at time t = 0) is not 0, but some value b0 ≥ 0. What is now the condition on R which ensures that no bit is discarded by the policer (in other words, that the flow is conformant) ? E XERCISE 1.28. Consider a variable capacity network node, with capacity curve M (t). Show that there is one maximum function S ∗ (t) such that for all 0 ≤ s ≤ t, we have M (t) − M (s) ≥ S ∗ (t − s) Show that S ∗ is super-additive. Conversely, if a function β is super-additive, show that there is a variable capacity network node, with capacity curve M (t), such that for all 0 ≤ s ≤ t, we have M (t) − M (s) ≥ S ∗ (t − s). Show that, with a notable exception, a shaper cannot be modeled as a variable capacity node. 1. Consider a packetized greedy shaper with shaping curve σ(t) = rt for t ≥ 0. E XERCISE 1.29. Assume that L(k) = kM where M is fixed. Assume that the input is given by R(t) = 10M for t > 0 and R(0) = 0. Compute the sequence R(i) (t) used in the representation of the output of the packetized greedy shaper, for i = 1, 2, 3, .... 2. Same question if σ(t) = (rt + 2M )1{ t > 0}. E XERCISE 1.30. Consider a source given by the function R(t) = B for t > 0 R(t) = 0 for t ≤ 0 Thus the flow consists of an instantaneous burst of B bits. 1. What is the minimum arrival curve for the flow ? 2. Assume that the flow is served in one node that offers a minimum service curve of the rate latency type, with rate r and latency ∆. What is the maximum delay for the last bit of the flow ? 3. We assume now that the flow goes through a series of two nodes, N1 and N2 , where Ni offers to the flow a minimum service curve of the rate latency type, with rate ri and latency ∆i , for i = 1, 2. What is the the maximum delay for the last bit of the flow through the series of two nodes ? 4. With the same assumption as in the previous item, call R1 (t) the function describing the flow at the output of node N1 (thus at the input of node N2 ). What is the worst case minimum arrival curve for R1 ? 5. We assume that we insert between N1 and N2 a “reformatter” S. The input to S is R1 (t). We call R1 (t) the output of S. Thus R1 (t) is now the input to N2 . The function of the “reformatter”S is to delay the flow R1 in order to output a flow R1 that is a delayed version of R. In other words, we must have R1 (t) = R(t − d) for some d. We assume that the reformatter S is optimal in the sense that it chooses the smallest possible d. In the worst case, what is this optimal value of d ?

1.11. EXERCISES

65

6. With the same assumptions as in the previous item, what is the worst case end-to-end delay through the series of nodes N1 , S, N2 ? Is the reformatter transparent ? E XERCISE 1.31. Let σ be a good function. Consider the concatenation of a bit-by-bit greedy shaper, with curve σ, and an L-packetizer. Assume that σ(0+ ) = 0. Consider only inputs that are L-packetized 1. Is this system a packetized shaper for σ ? 2. Is it a packetized shaper for σ + lmax ? 3. Is it a packetized greedy shaper for σ + lmax ? E XERCISE 1.32. Assume that σ is a good function and σ = σ0 + lu0 where u0 is the step function with a step at t = 0. Can we conclude that σ0 is sub-additive ? E XERCISE 1.33. Is the operator (P L ) upper-semi-continuous ? 1. Consider the concatenation of an L-packetizer and a network element with miniE XERCISE 1.34. mum service curve β and maximum service curve γ. Can we say that the combined system offer a minimum service curve (β(t) − lmax )+ and a maximum service curve γ, as in the case where the concatenation would be in the reverse order ? . 2. Consider the concatenation of a GPS node offering a guarantee λr1 , an L-packetizer, and a second GPS node offering a guarantee λr2 . Show that the combined system offers a rate-latency service curve lmax with rate R = min(r1 , r2 ) and latencyE = max(r . 1 ,r2 ) E XERCISE 1.35. Consider a node that offers to a flow R(t) a rate-latency service curve β = SR,L . Assume that R(t) is L-packetized, with packet arrival times called T1 , T2 , ... (and is left-continuous, as usual) Show that (R ⊗ β)(t) = minTi ∈[0,t] [R(Ti ) + β(t − Ti )] (and thus, the inf is attained). 1. Assume K connections, each with peak rate p, sustainable rate m and burst tolerE XERCISE 1.36. ance b, are offered to a trunk with constant service rate P and FIFO buffer of capacity X. Find the conditions on K for the system to be loss-free. 2. If Km = P , what is the condition on X for K connections to be accepted ? 3. What is the maximum number of connection if p = 2 Mb/s, m = 0.2 Mb/s, X = 10MBytes, b = 1Mbyte and P = 0.1, 1, 2 or 10 Mb/s ? 4. For a fixed buffer size X, draw the acceptance region when K and P are the variables. E XERCISE 1.37. Show the formulas giving the expressions for fB (R) and fB (α). 1. What is the effective bandwith for a connection with p = 2 Mb/s, m = 0.2 Mb/s, b = E XERCISE 1.38. 100 Kbytes when D = 1msec, 10 msec, 100 msec, 1s ? 2. Plot the effective bandwidth e as a function of the delay constraint in the general case of a connection with parameters p, m, b. 1. Compute the effective bandwidth for a mix of VBR connections 1, . . . , I. E XERCISE 1.39. 2. Show how the homogeneous case can be derived from your formula 3. Assume K connections, each with peak rate p, sustainable rate m and burst tolerance b, are offered to a trunk with constant service rate P and FIFO buffer of capacity X. Find the conditions on K for the system to be loss-free. 4. Assume that there are two classes of connections, with Ki connections in class i, i = 1, 2, offered to a trunk with constant service rate P and FIFO buffer of infinite capacity X. The connections are accepted as long as their queuing delay does not exceed some value D. Draw the acceptance region, that is, the set of (K1 , K2 ) that are accepted by CAC2. Is the acceptance region convex ? Is the complementary of the acceptance region in the positive orthant convex ? Does this generalize to more than two classes ?

66


C HAPTER 2

A PPLICATION OF N ETWORK C ALCULUS TO THE I NTERNET In this chapter we apply the concepts of Chapter 1 and explain the theoretical underpinnings of integrated and differentiated services. Integrated services define how reservations can be made for flows. We explain in detail how this framework was deeply influenced by GPS. In particular, we will see that it assumes that every router can be modeled as a node offering a minimum service curve that is a rate-latency function. We explain how this is used in a protocol such as RSVP. We also analyze the more efficient framework based on service curve scheduling. This allows us to address in a simple way the complex issue of schedulability. We explain the concept of Guaranteed Rate node, which corresponds to a service curve element, but with some differences, because it uses a max-plus approach instead of min-plus. We analyze the relation between the two approaches. Differentiated services differ radically, in that reservations are made per class of service, rather than per flow. We show how the bounding results in Chapter 1 can be applied to find delay and backlog bounds. We also introduce the “damper”, which is a way of enforcing a maximum service curve, and show how it can radically reduce the delay bounds.

2.1

GPS AND G UARANTEED R ATE N ODES

In this section we describe GPS and its derivatives; they form the basis on which the Internet guaranteed model was defined.

2.1.1

PACKET S CHEDULING

A guaranteed service network offers delay and throughput guarantees to flows, provided that the flows satisfy some arrival curve constraints (Section 2.2). This requires that network nodes implement some form of packet scheduling, also called service discipline. Packet scheduling is defined as the function that decides, at every buffer inside a network node, the service order for different packets. A simple form of packet scheduling is FIFO: packets are served in the order of arrival. The delay bound, and the required buffer, depend on the minimum arrival curve of the aggregate flow (Section 1.8 on page 53). If one flow sends a large amount of traffic, then the delay increases for all flows, and packet loss may occur. 67

CHAPTER 2. APPLICATION TO THE INTERNET

68

Thus FIFO scheduling requires that arrival curve constraints on all flows be strictly enforced at all points in the network. Also, with FIFO scheduling, the delay bound is the same for all flows. We study FIFO scheduling in more detail in Section 6. An alternative [25, 45] is to use per flow queuing, in order to (1) provide isolation to flows and (2) offer different guarantees. We consider first the ideal form of per flow queuing called “Generalized Processor Sharing” (GPS) [63], which was already mentioned in Chapter 1.

2.1.2

GPS AND A P RACTICAL I MPLEMENTATION (PGPS)

A GPS node serves several flows in parallel, and has a total output rate equal to c b/s. A flow i is allocated a given weight, say φi . Call Ri (t), Ri∗ (t) the input and output functions for flow i. The guarantee is that at any time t, the service rate offered to flow i is 0 is flow i has no backlog (namely, if Ri (t) = Ri∗ (t)), and otherwise is equal to φi φj c, where B(t) is the set of backlogged flows at time t. Thus j∈B(t)

Ri∗ (t) =

0

t

φi

j∈B(s) φj

1{i∈B(s)} ds

In the formula, we used the indicator function 1{expr} , which is equal to 1 if expr is true, and 0 otherwise. It follows immediately that the GPS node offers to flow i a service curve equal to λri c , with ri = φi Cφj . It j is shown in [64] that a better service curve can be obtained for every flow if we know some arrival curve properties for all flows; however the simple property is sufficient to understand the integrated service model. GPS satisfies the requirement of isolating flows and providing differentiated guarantees. We can compute the delay bound and buffer requirements for every flow if we know its arrival curve, using the results of Chapter 1. However, a GPS node is a theoretical concept, which is not really implementable, because it relies on a fluid model, and assumes that packets are infinitely divisible. How can we make a practical implementation of GPS ? One simple solution would be to use the virtual finish times as we did for the buffered leaky bucket controller in Section 1.7.3: for every packet we would compute its finish time θ under GPS, then at time θ present the packet to a multiplexer that serves packets at a rate c. Figure 2.1 (left) shows the finish times on an example. It also illustrates the main drawback that this method would have: at times 3 and 5, the multiplexer would be idle, whereas at time 6 it would have a burst of 5 packets to serve. In particular, such a scheduler would not be work conserving. This is what motivated researchers to find other practical implementations of GPS. We study here one such implementation of GPS, called packet by packet generalized processor sharing (PGPS) [63]. Other implementations of GPS are discussed in Section 2.1.3. PGPS emulates GPS as follows. There is one FIFO queue per flow. The scheduler handles packets one at a time, until it is fully transmitted, at the system rate c. For every packet, we compute the finish time that it would have under GPS (we call this the “GPS-finish-time”). Then, whenever a packet is finished transmitting, the next packet selected for transmission is the one with the earliest GPS-finish-time, among all packets present. Figure 2.1 shows one example. We see that, unlike the simple solution discussed earlier, PGPS is work conserving, but does so at the expense of maybe scheduling a packet before its finish time under GPS. We can quantify the difference between PGPS and GPS in the following proposition. In Section 2.1.3, we will see how to derive a service curve property. P ROPOSITION 2.1.1 ([63]). The finish time for PGPS is at most the finish time of GPS plus the total rate and L is the maximum packet size.

L c,

where c is

2.1. GPS AND GUARANTEED RATE NODES 0 1 2 3 4 5 6 7 8 9 1 0 1 1

69

tim e 0 1 2 3 4 5 6 7 8 9 1 0 1 1 flo w 0 1 2 3 4 5 A rriv a l D e p a rtu re

Figure 2.1: Scheduling with GPS (left) and PGPS (right). Flow 0 has weight 0.5, flows 1 to 5 have weight 0.1. All packets have the same transmission time equal to 1 time unit. P ROOF : Call D(n) the finish time of the nth packet for the aggregate input flow under PGPS, in the order of departure, and θ(n) under GPS. Call n0 the number of the packet that started the busy period in which packet n departs. Note that PGPS and GPS have the same busy periods, since if we observe only the aggregate flows, there is no difference between PGPS and GPS. There may be some packets that depart before packet n in PGPS, but that nonetheless have a later departure time under GPS. Call m0 ≥ n0 the largest packet number for which this occurs, if any; otherwise let m0 = n0 −1. In this proposition, we call l(m) the length in bits of packet m. Under PGPS, packet m0 started service at D(m0 )− l(mc 0 ) , which must be earlier than the arrival times of packets m = m0 +1, ..., n. Indeed, otherwise, by definition of PGPS, the PGPS scheduler would have scheduled packets m = m0 + 1, ..., n before packet m0 . Now let us observe the GPS system. Packets m = m0 + 1, ..., n depart no later than packet n, by definition of m0 ; they have arrived after D(m0 ) − l(mc 0 ) . By expressing the amount of service in the interval [D(m0 ) − l(mc 0 ) , θ(n)] we find thus n m=m0 +1

l(m0 ) l(m) ≤ c θ(n) − D(m0 ) + c

Now since packets m0 , ..., n are in the same busy period, we have n

m=m0 +1 l(m)

D(n) = D(m0 ) +

c

By combining the two equations above we find D(n) ≤ θ(n) + case where m0 ≥ n0 .

l(m0 ) c ,

which shows the proposition in the

If m0 = n0 − 1, then all packets n0 , ..., n depart before packet n under GPS and thus the same reasoning shows that n l(m) ≤ c (θ(n) − t0 ) m=n0

where t0 is the beginning of the busy period, and that D(n) = t0 + Thus D(n) ≤ θ(n) in that case.

n m=n0

c

l(m)


70

2.1.3

G UARANTEED R ATE (GR) N ODES AND THE M AX -P LUS A PPROACH

The service curve concept defined earlier can be approached from the dual point of view, which consists in studying the packet arrival and departure times instead of the functions R(t) (which count the bits arrived up to time t). This latter approach leads to max-plus algebra (which has the same properties as min-plus), is often more appropriate to account for details due to variable packet sizes, but works well only when the service curves are of the rate-latency type. It also useful when nodes cannot be assumed to be FIFO per flow, as may be the case with DiffServ (Section 2.4). GR also allows to show that many schedulers have the rate-latency service curve property. Indeed, a large number of practical implementations of GPS, other than PGSP, have been proposed in the literature; let us mention: virtual clock scheduling [49], packet by packet generalized processor sharing [63] and selfclocked fair queuing [40](see also [30]). For a thorough discussion of practical implementations of GPS, see [81, 30]). These implementations differ in their implementation complexity and in the bounds that can be obtained. It is shown in [32] that all of these implementations fit in the following framework, called “Guaranteed Rate”, which we define in now. We will also analyze how it relates to the min-plus approach. D EFINITION 2.1.1 (GR N ODE [32]). Consider a node that serves a flow. Packets are numbered in order of arrival. Let an ≥ 0, dn ≥ 0 be the arrival and departure times. We say that a node is the a guaranteed rate (GR) node for this flow, with rate r and delay e, if it guarantees that dn ≤ fn + e, where fn is defined by Equation (2.1). f0 = 0 (2.1) fn = max {an , fn−1 } + lrn for all n ≥ 1

The variables fn (“Guaranteed Rate Clocks”) can be interpreted as the departures times from a FIFO constant rate server, with rate r. The parameter e expresses how much the node deviates from it. Note however that a GR node need not be FIFO. A GR node is also called “Rate-Latency server”. Example: GPS. Consider an ideal GPS scheduler, which allocates a rate Ri = cφφi j to some flow i. It is a j GR node for flow i, with rate Ri and latency = 0 (the proof is left to the reader) D EFINITION 2.1.2 (O NE WAY D EVIATION OF A SCHEDULER FROM GPS). We say that S deviates from GPS by e if for all packet n the departure time satisfies dn ≤ gn + e

(2.2)

where gn is the departure time from a hypothetical GPS node that allocates a rate r = (assumed to be flow 1).

cφ1 j φj

to this flow

We interpret this definition as a comparison to a hypothetical GPS reference scheduler that would serve the same flows. T HEOREM 2.1.1. If a scheduler satisfies Equation (2.2), then it is GR with rate r and latency e. P ROOF :

gn ≤ fn and the rest is immediate.

Example: PGPS. Consider a PGPS scheduler, which allocates a rate Ri = L c,

cφi j φj

to some flow i. It is a GR

node for flow i, with rate Ri and latency where L is the maximum packet size (among all flows present at the scheduler) (this follows from Proposition 2.1.1).

2.1. GPS AND GUARANTEED RATE NODES

71

T HEOREM 2.1.2 (M AX -P LUS R EPRESENTATION OF GR). Consider a system where packets are numbered 1, 2, ... in order of arrival. Call an , dn the arrival and departure times for packet n, and ln the size of packet n. Define by convention d0 = 0. The system is a GR node with rate r and latency e if and only if for all n there is some k ∈ {1, ..., n} such that dn ≤ e + ak +

lk + ... + ln r

(2.3)

P ROOF : The recursion Equation (2.1) can be solved iteratively, using the same max-plus method as in the proof of Proposition 1.2.4. Define Anj = aj +

lj + ... + ln for 1 ≤ j ≤ n r

Then we obtain fn = max(Ann , Ann−1 , ..., An1 ) The rest follows immediately. Equation (2.3) is the dual of the service curve definition (Equation (1.9) on Page 71), with β(t) = r(t − e)+ . We now elucidate this relationship. T HEOREM 2.1.3 (E QUIVALENCE WITH SERVICE CURVE ). Consider a node with L-packetized input. 1. If the node guarantees a minimum service curve equal to the rate-latency function β = βr,v , and if it is FIFO, then it is a GR node with rate r and latency v. 2. Conversely, a GR node with rate r and latency e is the concatenation of a service curve element, with service curve equal to the rate-latency function βr,v , and an L-packetizer. If the GR node is FIFO, then so is the service curve element. The proof is long and is given at the end of this section. By applying Theorem 1.7.1, we obtain C OROLLARY 2.1.1. A GR node offers a minimum service curve βr,v+ lmax r

The service curve can be used to obtain backlog bounds. T HEOREM 2.1.4 (D ELAY B OUND ). For an α-smooth flow served in a (possibly non FIFO) GR node with rate r and latency e, the delay for any packet is bounded by sup[ t>0

P ROOF :

α(t) − t] + e r

By Theorem 2.1.2, for any fixed n, we can find a 1 ≤ j ≤ n such that fn = aj +

lj + ... + ln r

The delay for packet n is dn − an ≤ fn + e − an Define t = an − aj . By hypothesis

lj + ... + ln ≤ α(t+)

where α(t+) is the limit to the right of α at t. Thus dn − an ≤ −t + α(t+) Now supt>0 [ α(t) − t]. r − t] = supt≥0 [ r

α(t+) α(t+) + e ≤ sup[ − t] + e r r t≥0

(2.4)


72

C OMMENT: Note that Equation (2.4) is the horizontal deviation between the arrival curve α and the rate-latency service curve with rate r and latency e. Thus, for FIFO GR nodes, Theorem 2.1.4 follows from Theorem 2.1.2 and the fact that the packetizer can be ignored for delay computations. The information in Theorem 2.1.4 is that it also holds for non-FIFO nodes.

2.1.4

C ONCATENATION OF GR NODES

FIFO N ODES For GR nodes that are FIFO per flow, the concatenation result obtained with the service curve approach applies. (that are FIFO per flow) with rates rm T HEOREM 2.1.5. Specifically, the concatenation of M GR nodes and latencies em is GR with rate r = minm rm and latency e = i=1,...,n ei + i=1,...,n−1 Lmax ri , where Lmax is the maximum packet size for the flow. If rm = r for all m then the extra term is (M − 1) Lmax r ; it is due to packetizers. P ROOF : By Theorem 2.1.3–(2), we can decompose system i into a concatenation Si , Pi , where Si offers the service curve βri ,ei and Pi is a packetizer. Call S the concatenation

S1 , P1 , S2 , P2 , ..., Sn−1 , Pn−1 , Sn

By Theorem 2.1.3–(2), S is FIFO and provides the service curve βr,e . By Theorem 2.1.3–(1), it is GR with rate r and latency e. Now Pn does not affect the finish time of the last bit of every packet. Note that a slight change if the proof of the theorem showsthat the theorem is also valid if we replace Lmax by e = e = i=1,...,n ei + i=1,...,n−1 Lmax i=1,...,n ei + i=2,...,n ri . ri End-to-end Delay Bound. A bound on the end-to-end delay through a concatenation of GR nodes is thus D=

M m=1

vm + lmax

M −1 m=1

σ 1 + rm minm rm

(2.5)

which is the formula in [32]. It is a generalization of Equation (1.23) on Page 45. A Composite Node.We analyze in detail one specific example, which often arises in practice when modelling a router. We consider a composite node, made of two components. The former (“variable delay component”) imposes to packets a delay in the range [δmax − δ, δmax ]. The latter is FIFO and offers to its input the packet scale rate guarantee, with rate r and latency e. We show that, if the variable delay component is known to be FIFO, then we have a simple result. We first give the following lemma, which has some interest of its own. L EMMA 2.1.1 (VARIABLE D ELAY AS GR). Consider a node which is known to guarantee a delay ≤ δmax . The node need not be FIFO. Call lmin the minimum packet size. For any r > 0, the node is GR with latency + e = [δmax − lmin r ] and rate r. P ROOF : With the standard notation in this section, the hypothesis implies that dn ≤ an + δmax for all lmin n ≥ 1. Define fn by Equation (2.1). We have fn ≥ an + lrn ≥ an + lmin r , thus dn − fn ≤ δmax − r ≤ lmin + [δmax − r ] .

2.1. GPS AND GUARANTEED RATE NODES

73

T HEOREM 2.1.6. (Composite GR Node with FIFO Variable Delay Component) Consider the concatenation of two nodes. The former imposes to packets a delay ≤ δmax . The latter is a GR node with rate r and latency e. Both nodes are FIFO. The concatenation of the two nodes, in any order, is GR with rate r and latency e = e + δmax . + P ROOF : The former node is GR(r , e = [δmax − lmin r ] ) for any r > r. By Theorem 2.1.5 (and the note after it), the concatenation is GR(r, e + e + lmax r ). Let r go to ∞.

GR NODES THAT ARE NOT FIFO PER FLOW detail the composite node.

The concatenation result is no longer true. We study in

T HEOREM 2.1.7. Consider the concatenation of two nodes. The first imposes to packets a delay in the range [δmax − δ, δmax ]. The second is FIFO and offers the guaranteed rate clock service to its input, with rate r and latency e. The first node is not assumed to be FIFO, so the order of packet arrivals at the second node is not the order of packet arrivals at the first one. Assume that the fresh input is constrained by a continuous arrival curve α(·). The concatenation of the two nodes, in this order, offers to the fresh input the guaranteed rate clock service with rate r and latency e = e + δmax +

α(δ) − lmin r

The proof is given in the next section. Application: For α(t) = ρt + σ, we find e = e + δmax +

2.1.5

ρδ + σ − lmin r

P ROOFS

Proof of Theorem 2.1.3 Part 1: Consider a service curve element S. Assume to simplify the demonstration that the input and output functions R and R∗ are right-continuous. Consider the virtual system S 0 made of a bit-by-bit greedy shaper with shaping curve λr , followed by a constant bit-by-bit delay element. The bit-by-bit greedy shaper is a constant bit rate server, with rate r. Thus the last bit of packet n departs from it exactly at time fn , defined by Equation (2.1), thus the last bit of packet n leaves S 0 at d0n = fn + e. The output function of S 0 is R0 = R ⊗ βr,e . By hypothesis, R∗ ≥ R0 , and by the FIFO assumption, this shows that the delay in S is upper bounded by the delay in S . Thus dn ≤ fn + e. Part 2: Consider the virtual system S whose output S(t) is defined by if di−1 < t ≤ di then S(t) = min{R(t), max[L(i − 1), L(i) − r(di − t)]} See Figure 2.2 for an illustration. It follows immediately that R (t) = P L (S(t)). Also consider the virtual system S 0 whose output is S 0 (t) = (βr,v ⊗ R)(t) S 0 is the constant rate server, delayed by v. Our goal is now to show that S ≥ S 0 .

(2.6)


74

Call d0i the departure time of the last bit of packet i in S0 (see Figure 2.2 for an example with i = 2). Let u = d0i − di . The definition of GR node means that u ≥ 0. Now since S0 is a shifted constant rate server, we have: li if d0i − < s < d0i then S 0 (s) = L(i) − r(d0i − s) r Also d0i−1 ≤ d0i −

li r

thus S 0 (d0i − lri ) = L(i − 1) and if s ≤ d0i −

li then S 0 (s) ≤ L(i − 1) r

It follows that if di−1 + u < s < d0i then S 0 (s) ≤ max[L(i − 1), L(i) − r(d0i − s)]

(2.7)

Consider now some t ∈ (di−1 , di ] and let s = t + u. If S(t) = R(t), since R ≥ S 0 , we then obviously have S(t) ≥ S 0 (t). Else, from Equation (2.1), S(t) = max[L(i−1), L(i)−r(di −t)]. We have d0i −s = di −t and thus, combining with Equation (2.7), we derive that S 0 (s) ≤ S(t). Now s ≥ t, thus finally S 0 (t) ≤ S(t). One can also readily see that S is FIFO if di−1 ≤ di for all i.

R (t)

b its l3

R (t)

L (3 )

S (t)

l2

L (2 ) L (1 )

e r s lo p

l1 a 1

a 2

d 1

t

d 2

d

0

S 0(t)

2

Figure 2.2: Arrival and departure functions for GR node. The virtual system output is S(t). Proof of Theorem 2.1.7. We use the same notation and convention as in the proof of Theorem 7.5.3. We can also assume that all packet arrivals are distinct, using the same type of reduction. Fix some n ≥ 1; due to Theorem 2.1.2, it is sufficient to show that there is some k ∈ {1, ..., n} such that dn ≤ e2 + ak +

lk + ... + ln r

(2.8)

By hypothesis, there exists some j such that bj ≤ bn and dn ≤ bj + e +

B[bj , bn ] r

(2.9)

We cannot assume that j ≤ n; thus, define k as the oldest packet arrived in the interval [bj , bn ], in other words: k = inf{i ≥ 1 : bj ≤ bi ≤ bn }. Necessarily, we have now k ≤ n. Any packet that arrives at the second node in [bj , bn ] must have arrived at node 1 after or with packet k, and before bn . Thus B[bj , bn ] ≤ A[ak , bn ]. Now bn ≤ an + δ. Thus by Lemma 7.7.1 in this appendix: B[bj , bn ] ≤ A[ak , an ] + A(an , bn ] ≤ A[ak , an ] + α(δ) − lmin

2.2. THE INTEGRATED SERVICES MODEL OF THE IETF

75

Also, bj ≤ bk ≤ ak + δ and by Equation (2.9): dn ≤ ak + e + δ + α(δ) + A[ak , an ] − lmin which shows Equation (2.8).

2.2 2.2.1

T HE I NTEGRATED S ERVICES M ODEL OF THE IETF T HE G UARANTEED S ERVICE

The Internet supports different reservation principles. Two services are defined: the “guaranteed” service, and the “ controlled load” service. They differ in that the former provides real guarantees, while the latter provides only approximate guarantees. We outline the differences in the rest of this section. In both cases, the principle is based on “admission control”, which operates as follows. • In order to receive the guaranteed or controlled load service, a flow must first perform a reservation during a flow setup phase. • A flow must confirm to an arrival curve of the form α(t) = min(M + pt, rt + b), which is called the T-SPEC (see Section 1.2.2 on page13). The T-SPEC is declared during the reservation phase. • All routers along the path accept or reject the reservation. With the guaranteed service, routers accept the reservation only if they are able to provide a service curve guarantee and enough buffer for lossfree operation. The service curve is expressed during the reservation phase, as explained below. For the controlled load service, there is no strict definition of what accepting a reservation means. Most likely, it means that the router has an estimation module that says that, with good probability, the reservation can be accepted and little loss will occur; there is no service curve or delay guarantee. In the rest of this chapter we focus on the guaranteed service. Provision of the controlled load service relies on models with loss, which are discussed in Chapter 9.

2.2.2

T HE I NTEGRATED S ERVICES M ODEL FOR I NTERNET ROUTERS

The reservation phase assumes that all routers can export their characteristics using a very simple model. The model is based on the view that an integrated services router implements a practical approximation of GPS, such as PGPS, or more generally, a GR node. We have shown in Section 2.1.3 that the service curve offered to a flow by a router implementing GR is a rate-latency function, with rate R and latency T connected by the relationship C +D (2.10) T = R with C = the maximum packet size for the flow and D = Lc , where L is the maximum packet size in the router across all flows, and c the total rate of the scheduler. This is the model defined for an Internet node [75]. FACT 2.2.1. The Integrated Services model for a router is that the service curve offered to a flow is always a rate-latency function, with parameters related by a relation of the form (2.10). The values of C and D depend on the specific implementation of a router, see Corollary 2.1.1 in the case of GR nodes. Note that a router does not necessarily implement a scheduling method that approximates GPS. In fact, we discuss in Section 2.3 a family of schedulers that has many advantages above GPS. If a router implements a method that largely differs from GPS, then we must find a service curve that lower-bounds the best service curve guarantee offered by the router. In some cases, this may mean loosing important


76

information about the router. For example, it is not possible to implement a network offering constant delay to flows by means of a system like SCED+, discussed in Section 2.4.3, with the Integrated Services router model.

2.2.3

R ESERVATION S ETUP WITH RSVP

Consider a flow defined by TSPEC (M, p, r, b), that traverses nodes 1, . . . , N . Usually, nodes 1 and N are end-systems while nodes n for 1 < n < N are routers. The Integrated Services model assumes that node n on the path of the flow offers a rate latency service curve βRn ,Tn , and further assumes that Tn has the form Tn =

Cn + Dn R

where Cn and Dn are constants that depend on the characteristics of node n. The reservation is actually put in place by means of a flow setup procedure such as the resource reservation protocol (RSVP). At the end of the procedure, node n on the path has allocated to the flow a value Rn ≥ r. This is equivalent to allocating a service curve βRn ,Tn . From Theorem 1.4.6 on page 28, the end-to-end service curve offered to the flow is the rate-latency function with rate R and latency T given by R = minn=1...N Rn N Cn T = n=1 Rn + Dn Let Ctot =

N

n=1 Cn

and Dtot =

N

n=1 Dn .

We can re-write the last equation as

C Sn T = tot + Dtot − R N

(2.11)

n=1

with Sn = Cn

1 1 − R Rn

(2.12)

The term Sn is called the “local slack” term at node n. From Proposition 1.4.1 we deduce immediately: P ROPOSITION 2.2.1. If R ≥ r, the bound on the end-to-end delay, under the conditions described above is

N b − M p − R + M + Ctot + Dtot − + Sn (2.13) R p−r R n=1

We can now describe the reservation setup with RSVP. Some details of flow setup with RSVP are illustrated on Figure 2.3. It shows that two RSVP flows are involved: an advertisement (PATH) flow and a reservation (RESV) flow. We describe first the point-to-point case. • A PATH message is sent by the source; it contains the T-SPEC of the flow (source T-SPEC), which is not modified in transit, and another field, the ADSPEC, which is accumulated along the path. At a destination, the ADSPEC field contains, among others, the values of Ctot , Dtot used in Equation 2.13. PATH messages do not cause any reservation to be made. • RESV messages are sent by the destination and cause the actual reservations to be made. They follow the reverse path marked by PATH messages. The RESV message contains a value, R , (as part of the so-called R-SPEC), which is a lower bound on the rate parameters Rn that routers along the path will have to reserve. The value of R is determined by the destination based on the end-to-end delay objective dobj , following the procedure described below. It is normally not changed by the intermediate nodes.

2.2. THE INTEGRATED SERVICES MODEL OF THE IETF

77

R e ce iv e r B R1 R2 1 . p a th m e ssa g e 2 . p a th m e ssa g e TSP EC = 3 . p a th m e ssa g e Se n d e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K Se n d e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K A d Sp e c= () 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K A d Sp e c= (1 0 .2 k b , 0 .0 5 s) A d Sp e c= (5 1 .2 , 0 .1 ) Sender A

4 . B re q u e sts g u a ra n te e d Q o S re se rv a tio n w ith d e la y v a ria tio n 0 .6 s; B re se rv e s 6 2 2 k b /s 5 . re sv m e ssa g e 6 . re sv m e ssa g e R e ce iv e r TSP EC = 7 . re sv m e ssa g e R e ce iv e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R e ce iv e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R -SP EC = (6 2 2 k b /s) 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R -SP EC = (6 2 2 k b /s) R -SP EC = (6 2 2 k b /s)

Figure 2.3: Setup of Reservations, showing the PATH and RESV flows Define function f by f (R ) :=

b−M R

p − R p−r

+ +

M + Ctot + Dtot R

In other words, f is the function that defines the end-to-end delay bound, assuming all nodes along the path would reserve Rn = R . The destination computes R as the smallest value ≥ r for which f (R ) ≤ dobj . Such a value exists only if Dtot < dobj . In the figure, the destination requires a delay variation objective of 600 ms, which imposes a minimum value of R =622 kb/s. The value of R is sent to the next upstream node in the R-SPEC field of the PATH message. The intermediate nodes do not know the complete values Ctot and Dtot , nor do they know the total delay variation objective. Consider the simple case where all intermediate nodes are true PGPS schedulers. Node n simply checks whether it is able to reserve Rn = R to the flow; this involves verifying that the sum of reserved rates is less than the scheduler total rate, and that there is enough buffer available (see below). If so, it passes the RESV message upstream, up to the destination if all intermediate nodes accept the reservation. If the reservation is rejected, then the node discards it and normally informs the source. In this simple case, all nodes should set their rate to Rn = R thus R = R , and Equation (2.13) guarantees that the end-to-end delay bound is guaranteed. In practice, there is a small additional element (use of the slack term), due to the fact that the designers of RSVP also wanted to support other schedulers. It works as follows. There is another term in the R-SPEC, called the slack term. Its use is illustrated on Figure 2.4. In the figure, we see that the end-to-end delay variation requirement, set by the destination, is 1000 ms. In that case, the destination reserves the minimum rate, namely, 512 kb/s. Even so, the delay variation objective Dobj is larger than the bound Dmax given by Formula (2.13). The difference Dobj − Dmax is written in the slack term S and passed to the upstream node in the RESV message. The upstream node is not able to compute Formula (2.13) because it does not have the value of the end-to-end parameters. However, it can use the slack term to increase its internal delay objective, on top of what it had advertised. For example, a guaranteed rate node may increase its value of v (Theorem 2.1.2) and thus reduce the internal resources required to perform the reservation. The figure shows that R1 reduces the slack term by 100 ms. This is equivalent to increasing the Dtot parameter by 100ms, but without modifying the advertised Dtot .


78

R e ce iv e r R1 R2 B 1 . p a th m e ssa g e 2 . p a th m e ssa g e TSP EC = 3 . p a th m e ssa g e Se n d e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K Se n d e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K A d Sp e c= () 2 K ,1 0 M b /s,5 1 2 k b /s,3 2 K A d Sp e c= (1 0 .2 s/k b /s, A d Sp e c= (5 1 .2 , 0 .1 ) 0 .0 5 s) Sender A

4 . B re q u e sts g u a ra n te e d Q o S re se rv a tio n w ith d e la y v a ria tio n 1 .0 s; B re se rv e s 5 1 2 k b /s 5 . re sv m e ssa g e 6 . re sv m e ssa g e R e ce iv e r TSP EC = 6 . re sv m e ssa g e R e ce iv e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R e ce iv e r TSP EC = 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R -SP EC = (5 1 2 k b /s , S= 2 K ,1 0 M b /s,5 1 2 k b /s,2 4 K R -SP EC = (5 1 2 k b /s , S= 0 .2 8 8 s) R -SP EC = (5 1 2 k b /s, 0 .2 8 8 s) S= 0 .1 8 8 s)

Figure 2.4: Use of the slack term The delays considered here are the total (fixed plus variable) delays. RSVP also contains a field used for advertising the fixed delay part, which can be used to compute the end-to-end fixed delay. The variable part of the delay (called delay jitter) is then obtained by subtraction.

2.2.4

A F LOW S ETUP A LGORITHM

There are many different ways for nodes to decide which parameter they should allocate. We present here one possible algorithm. A destination computes the worst case delay variation, obtained if all nodes reserve the sustainable rate r. If the resulting delay variation is acceptable, then the destination sets R = r and the resulting slack may be used by intermediate nodes to add a local delay on top of their advertised delay variation defined by C and D. Otherwise, the destination sets R to the minimum value Rmin that supports the end-to-end delay variation objective and sets the slack to 0. As a result, all nodes along the path have to reserve Rmin . As in the previous cases, nodes may allocate a rate larger than the value of R they pass upstream, as a means to reduce their buffer requirement. D EFINITION 2.2.1 (A F LOW S ETUP A LGORITHM ).

• At a destination system I, compute

Dmax = fT (r) +

Ctot + Dtot r

If Dobj > Dmax then assign to the flow a rate RI = r and an additional delay variation dI ≤ Dobj − Dmax ; set SI = Dobj − Dmax − dI and send reservation request RI , SI to station I − 1. tot Else (Dobj ≤ Dmax ) find the minimum Rmin such that fT (Rmin ) + RCmin ≤ Dobj − Dtot , if it exists. Send reservation request RI = Rmin , SI = 0 to station I − 1. If Rmin does not exist, reject the reservation or increase the delay variation objective Dobj . • At an intermediate system i: receive from i + 1 a reservation request Ri+1 , Si+1 . If Si = 0, then perform reservation for rate Ri+1 and if successful, send reservation request Ri = Ri+1 , Si = 0 to station i − 1. Else (Si > 0), perform a reservation for rate Ri+1 with some additional delay variation di ≤ Si+1 . if successful, send reservation request Ri = Ri+1 , Si = Si+1 − di to station i − 1.

2.3. SCHEDULABILITY

79

The algorithm ensures a constant reservation rate. It is easy to check that the end to end delay variation is bounded by Dobj .

2.2.5

M ULTICAST F LOWS

Consider now a multicast situation. A source S sends to a number of destinations, along a multicast tree. PATH messages are forwarded along the tree, they are duplicated at splitting points; at the same points, RESV messages are merged. Consider such a point, call it node i, and assume it receives reservation requests , S and R , S . The node performs reservations for the same T-SPEC but with respective parameters Rin in in in internally, using the semantics of algorithm 3. Then it has to merge the reservation requests it will send to node i − 1. Merging uses the following rules: R-SPEC M ERGING RULES

The merged reservation R, S is given by R = max(R , R ) S = min(S , S )

Let us consider now a tree where algorithm 3 is applied. We want to show that the end-to-end delay bounds at all destinations are respected. The rate along the path from a destination to a source cannot decrease with this algorithm. Thus the minimum rate along the tree towards the destination is the rate set at the destination, which proves the result. A few more features of RSVP are: • states in nodes need to be refreshed; if they are not refreshed, the reservation is released (“soft states”). • routing is not coordinated with the reservation of the flow We have so far looked only at the delay constraints. Buffer requirements can be computed using the values in Proposition 1.4.1.

2.2.6

F LOW S ETUP WITH ATM

With ATM, there are the following differences: • The path is determined at the flow setup time only. Different connections may follow different routes depending on their requirements, and once setup, a connection always uses the same path. • With standard ATM signaling, connection setup is initiated at the source and is confirmed by the destination and all intermediate systems.

2.3

S CHEDULABILITY

So far, we have considered one flow in isolation and assumed that a node is able to offer some scheduling, or service curve guarantee. In this section we address the global problem of resource allocation. When a node performs a reservation, it is necessary to check whether local resources are sufficient. In general, the method for this consists in breaking the node down into a network of building blocks such as schedulers, shapers, and delay elements. There are mainly two resources to account for: bit rate (called “bandwidth”) and buffer. The main difficulty is the allocation of bit rate. Following [36], we will see in this


80

section that allocating a rate amounts to allocating a service curve. It is also equivalent to the concept of schedulability. Consider the simple case of a PGPS scheduler, with outgoing rate C. If we want to allocate rate ri to flow i, for every i, then we can allocate to flow i the GPS weight φi = rCi . Assume that

ri ≤ C

(2.14)

i

Then we know from Proposition 2.1.1 and Corollary 2.1.1 that every flow i is guaranteed the rate-latency L . In other words, the schedulability condition for PGPS is simply service curve with rate ri and latency C Equation (2.14). However, we will see now that a schedulability conditions are not always as simple. Note also that the end-to-end delay depends not only on the service curve allocated to the flow, but also on its arrival curve constraints. Many schedulers have been proposed, and some of them do not fit in the GR framework. The most general framework in the context of guaranteed service is given by SCED (Service Curve Earliest Deadline first) [36],which we describe now. We give the theory for constant size packets and slotted time; some aspects of the general theory for variable length packets are known [11], some others remain to be done. We assume without loss of generality that every packet is of size 1 data unit.

2.3.1

EDF S CHEDULERS

As the name indicates, SCED is based on the concept of Earliest Deadline First (EDF) scheduler. An EDF scheduler assigns a deadline Din to the nth packet of flow i, according to some method. We assume that deadlines are wide-sense increasing within a flow. At every time slot, the scheduler picks at one of the packets with the smallest deadline among all packets present. There is a wide variety of methods for computing deadlines. The “delay based” schedulers [55] set Din = An + di where An is the arrival time for the nth packet for flow i, and di is the delay budget allocated to flow i. If di is independent of i, then we have a FIFO scheduler. We will see that those are special cases of SCED, which we view as a very general method for computing deadlines. An EDF scheduler is work conserving, that is, it cannot be idle if there is at least one packet present in the system. A consequence of this is that packets from different flows are not necessarily served in the order of their deadlines. Consider for example a delay based scheduler, and assume that flow 1 has a lrage delay budget d1 , while flow 2 has a small delay budget d2 . It may be that a packet of flow 1 arriving at t1 is served before a packet of flow 2 arriving at t2 , even though the deadline of packet 1, t1 + d1 is larger than the deadline of packet 2. We will now derive a general schedulability criterion for EDF schedulers. Call Ri (t), t ∈ N, the arrival function for flow i. Call Zi (t) the number of packets of flow i that have deadlines ≤ t. For example, for a delay based scheduler, Zi (t) = Ri (t − di ). The following is a modified version of [11]. P ROPOSITION 2.3.1. Consider an EDF scheduler with I flows and outgoing rate C. A necessary condition for all packets to be served within their deadlines is for all s ≤ t :

I

Zi (t) − Ri (s) ≤ C(t − s)

(2.15)

i=1

A sufficient condition is for all s ≤ t :

I [Zi (t) − Ri (s)]+ ≤ C(t − s) i=1

(2.16)

2.3. SCHEDULABILITY

81

P ROOF : We first prove the necessary condition. Call Ri the output for flow i. Since the scheduler is I I work conserving, we have i=1 Ri = λC ⊗ ( i=1 Ri ). Now Ri ≥ Zi by hypothesis. Thus I

Zi (t) ≤ inf C(t − s) + s∈[0,t]

i=1

I

Ri (s)

i=1

which is equivalent to Equation (2.15) Now we prove the sufficient condition, by contradiction. Assume that at some t a packet with deadline t is not yet served. In time slot t, the packet served has a deadline ≤ t, otherwise our packet would have been chosen instead. Define s0 such that the time interval [s0 + 1, t] is the maximum time interval ending at t that is within a busy period and for which all packets served have deadlines ≤ t. Now call S the set of flows that have a packet with deadline ≤ t present in the system at some point in the interval [s0 + 1, t]. We show that if if i ∈ S then Ri (s0 ) = Ri (s0 )

(2.17)

that is, flow i is not backlogged at the end of time slot s0 . Indeed, if s0 + 1 is the beginning of the busy period, then the property is true for any flow. Otherwise, we proceed by contradiction. Assume that i ∈ S and that i would have some backlog at the end of time slot s0 . At time s0 some packet with deadline > t was served; thus the deadline of all packets remaining in the queue at the end of time slot s0 must have a deadline > t. Since deadlines are assumed wide-sense increasing within a flow, all deadlines of flow i packets that are in the queue at time s0 , or will arrive later, have deadline > t, which contradicts that i ∈ S. Further, it follows from the last argument that if i ∈ S, then all packets served before or at t must have a deadline ≤ t. Thus if i ∈ S then Ri (t) ≤ Zi (t) Now since there is at least one packet with deadline ≤ t not served at t, the previous inequality is strict for at least one i in S. Thus Ri (t) < Zi (t) (2.18) i∈S

i∈S

Observe that all packets served in [s0 + 1, t] must be from flows in S. Thus I

(Ri (t) − Ri (s0 )) =

i=1

(Ri (t) − Ri (s0 ))

i∈S

Combining with Equation (2.17) and Equation (2.18) gives I

(Ri (t) − Ri (s0 ))
h−1 the worst case delay does grow to infinity [41]. In some cases the network may be unbounded; in some other cases (such as the unidirectional ring, there is always a finite bound for all ν < 1. This issue is discussed in Chapter 6, where we we find better bounds, at the expense of more restrictions on the routes and the rates. Such restrictions do not fit with the differentiated services framework. Note also that, for feed-forward networks, we know that there are finite bounds for ν < 1. However we show now that the 1 condition ν < h−1 is the best that can be obtained, in some sense.

P ROPOSITION 2.4.1. [4, 14] With the assumptions of Theorem 2.4.1, if ν > is a network in which the worst case delay is at least D .

1 h−1 ,

then for any D > 0, there

In other words, the worst case queuing delay can be made arbitrarily large; thus if we want to go beyond Theorem 2.4.1, any bound for differentiated services must depend on the network topology or size, not only on the utilization factor and the number of hops. P ROOF : We build a family of networks, out of which, for any D , we can exhibit an example where the queuing delay is at least D . The thinking behind the construction is as follows. All flows are low priority flows. We create a hierarchical network, where at the first level of the hierarchy we choose one “flow” for which its first packet happens to encounter just one packet of every other flow whose route it intersects, while its next packet does not encounter any queue at all. This causes the first two packets of the chosen flow to come back-to-back after several hops. We then construct the second level of the hierarchy by taking a new flow and making sure that its first packet encounters two back-to-back packets of each flow whose routes it intersects, where the two back-to-back packet bursts of all these flows come from the output of a sufficient number of networks constructed as described at the first level of the hierarchy. Repeating this process recursively sufficient number of times, for any chosen delay value D we can create deep enough hierarchy so that the queuing

2.4. APPLICATION TO DIFFERENTIATED SERVICES

91

delay of the first packet of some flow encounters a queuing delay more than D (because it encounters a large enough back-to-back burst of packets of every other flow constructed in the previous iteration), while the second packet does not suffer any queuing delay at all. We now describe in detail how to construct such a hierarchical network (which is really a family of networks) such that utilization factor of any link does not exceed a given factor ν, and no flow traverses more than h hops. Now let us describe the networks in detail. We consider a family of networks with a single traffic class and constant rate links, all with same bit rate C. The network is assumed to be made of infinitely fast switches, with one output buffer per link. Assume that sources are all leaky bucket constrained, but are served in an aggregate manner, first in first out. Leaky bucket constraints are implemented at the network entry; after that point, all flows are aggregated. Without loss of generality, we also assume that propagation delays can be set to 0; this is because we focus only on queuing delays. As a simplification, in this network, we also assume that all packets have a unit size. We show that for any fixed, but arbitrary delay budget D, we can build a network of that family where the worst case queueing delay is larger than D, while each flow traverses at most a specified number of hops. A network in our family is called N (h, ν, J) and has three parameters: h (maximum hop count for any 1 flow), ν (utilization factor) and J (recursion depth). We focus on the cases where h ≥ 3 and h−1 < ν < 1, which implies that we can always find some integer k such that ν>

1 kh + 1 h − 1 kh − 1

(2.24)

Network N (h, ν, J) is illustrated in Figures 2.8 and 2.9; it is a collection of identical building blocks, arranged in a tree structure of depth J. Every building block has one internal source of traffic (called “transit traffic”), kh(h − 1) inputs (called the “building block inputs”), kh(h − 1) data sinks, h − 1 internal nodes, and one output. Each of the h − 1 internal nodes receives traffic from kh building block inputs plus it receives transit traffic from the previous internal node, with the exception of the first one which is fed by the internal source. After traversing one internal node, traffic from the building block inputs dies in a data sink. In contrast, transit traffic is fed to the next internal node, except for the last one which feeds the building block output (Figure 2.8). Figure 2.9 illustrates that our network has the structure of a complete tree, with depth J. The building blocks are organized in levels j = 1, ..., J. Each of the inputs of a level j building block (j ≥ 2) is fed by the output of one level j − 1 building block. The inputs of level 1 building blocks are data sources. The output of one j − 1 building block feeds exactly one level j building block input. At level J, there is exactly one building block, thus at level J − 1 there are kh(h − 1) building blocks, and at νC and burst level 1 there are (kh(h − 1))J−1 building blocks. All data sources have the same rate r = kh+1 tolerance b = 1 packet. In the rest of this section we take as a time unit the transmission time for one packet, so that C = 1. Thus any source may transmit one packet every θ = kh+1 time units. Note that a source ν may refrain from sending packets, which is actually what causes the large delay jitter. The utilization factor on every link is ν, and every flow uses 1 or h hops. Now consider the following scenario. Consider some arbitrary level 1 building block. At time t0 , assume that a packet fully arrives at each of the building block inputs of level 1, and at time t0 + 1, let a packet fully arrive from each data source inside every level 1 building block (this is the first transit packet). The first transit packet is delayed by hk − 1 time units in the first internal node. Just one time unit before this packet leaves the first queue, let one packet fully arrive at each input of the second internal node. Our first transit packet will be delayed again by hk − 1 time units. If we repeat the scenario along all internal nodes inside the building block, we see that the first transit packet is delayed by (h − 1)(hk − 1) time units. Now from Equation (2.24), θ < (h − 1)(hk − 1), so it is possible for the data source to send a second transit packet at time (h − 1)(hk − 1). Let all sources mentioned so far be idle, except for the emissions already described. The second transit packet will catch up to the first one, so the output of any level 1 building block is a burst of two back-to-back packets. We can choose t0 arbitrarily, so we have a mechanism for generating bursts of 2 packets.


92

b u ffe r

m u ltip le x e r

1 d a ta so u rc e

d e m u ltip le x e r

(h -1 ) k h in p u ts 1 o u tp u t h -1 in te rn a l n o d e s

(h -1 ) k h d a ta s in k s

Figure 2.8: The internal node (top) and the building block (bottom) used in our network example.

le v e l J - 2

le v e l J -1 le v e l J Figure 2.9: The network made of building blocks from Figure 2.8


93

Now we can iterate the scenario and use the same construction at level 2. The level-2 data source sends exactly three packets, spaced by θ. Since the internal node receives hk bursts of two packets originating from level 1, a judicious choice of the level 1 starting time lets the first level 2 transit packet find a queue of 2hk − 1 packets in the first internal node. With the same construction as in level 1, we end up with a total queuing delay of (h − 1)(2hk − 1) > 2(h − 1)(hk − 1) > 2θ for that packet. Now this delay is more than 2θ, and the first three level-2 transit packets are delayed by the same set of non-transit packets; as a result, the second and third level-2 transit packets will eventually catch up to the first one and the output of a level 2 block is a burst of three packets. This procedure easily generalizes to all levels up to J. In particular, the first transit packet at level J has an end-to-end delay of at least Jθ. Since all sources become idle after some time, we can easily create a last level J transit packet that finds an empty network and thus a zero queuing delay. Thus there are two packets in network N (h, ν, J), with one packet having a delay larger than Jθ, and the other packet has zero delay. This establishes that a bound on queuing delay, and thus on delay variation in network N (h, ν, J) has to be at least as large as Jθ.

2.4.3

B OUNDS FOR AGGREGATE S CHEDULING WITH DAMPERS

At the expense of some protocol complexity, the previous bounds can be improved without losing the feature of aggregate scheduling. It is even possible to avoid bound explosions at all, using the concepts of damper. Consider an EDF scheduler (for example a SCED scheduler) and assume that every packet sent on the outgoing link carries a field with the difference d between its deadline and its actual emission time, if it is positive, and 0 otherwise. A damper is a regulator in the next downstream node that picks for the packet an eligibility time that lies in the interval [a + d − ∆, a + d], where ∆ is a constant of the damper, and a is the arrival time of the packet in the node where the damper resides. We call ∆ the “damping tolerance”. The packet is then withheld until its eligibility time [80, 20], see Figure 2.10. In addition, we assume that the damper operates in a FIFO manner; this means that the sequence of eligibility times for consecutive packets is wide-sense increasing. Unlike the scheduler, the damper does not exist in isolation. It is associated with the next scheduler on the path of a packet. Its effect is to forbid scheduling the packet before the eligibility time chosen for the packet. Consider Figure 2.10. Scheduler m works as follows. When it has an opportunity to send a packet, say at time t, it picks a packet with the earliest deadline, among all packets that are present in node N , and whose eligibility date is ≥ t. The timing information d shown in the figure is carried in a packet header, either as a link layer header information, or as an IP hop by hop header extension. At the end of a path, we assume that there is no damper at the destination node. The following proposition is obvious, but important, and is given without proof. P ROPOSITION 2.4.2. Consider the combination S of a scheduler and its associated damper. If all packets are served by the scheduler before or at their deadlines, then S provides a bound on delay variation equal to ∆. It is possible to let ∆ = 0, in which case the delay is constant for all packets. A bound on the end-to-end delay variation is then the delay bound at the last scheduler using the combination of a scheduler and a damper (this is called “jitter EDD” in [80]). In practice, we consider ∆ > 0 for two reasons. Firstly, it is impractical to assume that we can write the field d with absolute accuracy. Secondly, having some slack in the delay variation objective provides better performance to low priority traffic [20]. There is no complicated feasibility condition for a damper, as there is for schedulers. The operation of a damper is always possible, as long as there is enough buffer. P ROPOSITION 2.4.3 (B UFFER REQUIREMENT FOR A DAMPER ). If all packets are served by the scheduler


94

S w it c h in g F a b r ic S c h e d u le r k

R o u t e r

F r o m o t h e r r o u t e r

T o o t h e r r o u t e r

D a m p e r n

R o u t e r M

S c h e d u le r l

D e p a r t u r e f r o m

S c h e d u le r m

N

D a m p e r l D e a d lin e a t l

l

d P a c k e t s e n t f r o m M t o N d

A r r iv a l a t n o d e N

a

a + d - d

a + d

E lig ib ilit y t im e p ic k e d b y d a m p e r l

Figure 2.10: Dampers in a differentiated services context. The model shown here assumes that routers are made of infinitely fast switching fabrics and output schedulers. There is one logical damper for each upstream scheduler. The damper decides when an arriving packet becomes visible in the node.


95

before or at their deadlines, then the buffer requirement at the associated damper is bounded by the buffer requirement at the scheduler. P ROOF : Call R( t) the total input to the scheduler, and R (t) the amount of data with deadline ≤ t. Call R∗ (t) the input to the damper, we have R∗ (t) ≤ R(t). Packets do not stay in the damper longer than until their deadline in the scheduler, thus the output R1 (t) of the damper satisfies R1 (t) ≥ R (t). The buffer requirement at the scheduler at time t is R(t) − R (t); at the damper it is R∗ (t) − R1 (t) ≥ R(t) − R (t). T HEOREM 2.4.3 (D ELAY AND BACKLOG BOUNDS WITH DAMPERS ). Take the same assumptions as in Theorem 2.4.1, we assume that every scheduler m that is not an exit point is associated with a damper in the next downstream node, with damping tolerance ∆m . Let ∆ be a bound on all ∆m . If ν ≤ 1, then a bound on the end-to-end delay jitter for low delay traffic is D = e + (h − 1)∆(1 + ν) + τ ν A bound on the queuing delay at any scheduler is D0 = e + ν[τ + (h − 1)∆] The buffer required at scheduler m, for serving low delay traffic without loss is bounded by Breq = rm D0 A bound on the buffer required at damper m is the same as the buffer required at scheduler m. P ROOF : The variable part of the delay between the input of a scheduler and the input of the next one is bounded by ∆. Now let us examine the last scheduler, say m, on the path of a packet. The delay between a source for a flow i m and scheduler m is a constant plus a variable part bounded by (h − 1)∆. Thus an arrival curve for the aggregate low-delay traffic arriving at scheduler m is α2 (t) = νrm (t + τ + (h − 1)∆) By applying Theorem 1.4.2, a delay bound at scheduler m is given by D2 = E + uν[τ + (h − 1)∆] A bound on end-to-end delay variation is (h − 1)∆ + D2 , which is the required formula. The derivation of the backlog bound is similar to that in Theorem 2.4.1. The benefit of dampers is obvious: there is no explosion to the bound, it is finite (and small if ∆ is small) for any utilization factor up to 1 (see Figure 2.11). Furthermore, the bound is dominated by h∆, across the whole range of utilization factors up to 1. A key factor in obtaining little delay variation is to have a small damping tolerance δ. There is a relation between a damper and a maximum service curve. Consider the combination of a scheduler with minimum service curve β and its associate damper with damping tolerance ∆. Call p the fixed delay on the link between the two. It follows immediately that the combination offers the maximum service curve β ⊗ δp−∆ and the minimum service curve β ⊗ δp . Thus a damper may be viewed as a way to implement maximum service curve guarantees. This is explored in detail in [20].

2.4.4

S TATIC E ARLIEST T IME F IRST (SETF)

A simpler alternative to the of dampers is proposed by Z.-L. Zhang et al under the name of Static Earliest Time First (SETF) [84].


96

0.5 0.4 0.3 0.2 0.1

0.2

0.4

0.6

0.8

Figure 2.11: The bound D (in seconds) in Theorem 2.4.3 the same parameters as Figure 2.7, for a damping tolerance ∆ = 5 ms per damper, and Cm = +∞ (thick line). The figure also shows the two curves of Figure 2.7, for comparison. The bound is very close to h∆ = 0.05s, for all utilization factors up to 1. A SSUMPTIONS

We take the same assumptions as with Theorem 2.4.1, with the following differences.

• At network access, packets are stamped with their time of arrival. At any node, they are served within the EF aggregate at one node in order of time stamps. Thus we assume that nodes offer a GR guarantee to the EF aggregate, as defined by Equation (2.1) or Equation (2.3), but where packets are numbered in order of time stamps (i.e. their order at the network access, not at this node). T HEOREM 2.4.4. If the time stamps have infinite precision, for all ν < 1, the end-to-end delay variation for the EF aggregate is bounded by 1 − (1 − ν)h D = (e + τ ) ν(1 − ν)h−1

P ROOF : The proof is similar to the proof of Theorem 2.4.1. Call Dk the least bound, assuming it exists, on the end-to-end delay after k hops, k ≤ h. Consider a tagged packet, with label n, and call dk its delay in k hops. Consider the node m that is the hth hop for this packet. Apply Equation (2.3): there is some label k ≤ n such that lk + ... + ln (2.25) dn ≤ e + ak + r where aj and dj are the arrival and departure times at node m of the packet labeled j, and lj its length in bits. Now packets k to n must have arrived at the network access before an − dk and after am − DHh−1 . Thus lk + ... + ln ≤ α(an − am − dk + Dh−1 ) where α is an arrival curve at network access for the traffic that will flow through node m. We have α(t) ≤ rm (νt + τ ). By Equation (2.4), the delay dn − an for our tagged packet is bounded by α(t − dk + Dh−1 ) − t = e + τ + ν(Dh−1 − dk ) e + sup rm t≥0 thus dk+1 ≤ dk + e + τ + ν(Dh−1 − dk )

2.5. BIBLIOGRAPHIC NOTES

97

The above inequation can be solved iteratively for dk as a function of Dh−1 ; then take k = h − 1 and assume the tagged packet is one that achieves the worst case k-hop delay, thus Dh−1 = dh−1 which gives an inequality for Dh−1 ; last, take k = h and obtain the end-to-end delay bound as desired. C OMMENTS : The bound is finite for all values of the utilization factor ν < 1, unlike the end-to-end bound in Theorem 2.4.1. Note that for small values of ν, the two bounds are equivalent. We have assumed here infinite precision about the arrival time stamped in every packet. In practice, the timestamp is written with some finite precision; in that case, Zhang [84] finds a bound which lies between Theorem 2.4.1 and Theorem 2.4.4 (at the limit, with null precision, the bound is exactly Theorem 2.4.4).

2.5

B IBLIOGRAPHIC N OTES

The delay bound for EF in Theorem 2.4.2 was originally found in [14], but neglecting the Lmax term; a formula that accounts for Lmax was found in [43]. Bounds that account for statistical multiplexing can be found in [58].

2.6

E XERCISES

E XERCISE 2.1. Consider a guaranteed rate scheduler, with rate R and delay v, that receives a packet flow with cumulative packet length L. The (packetized) scheduler output is fed into a constant bit rate trunk with rate c > R and propagation delay T . 1. Find a minimum service curve for the complete system. 2. Assume the flow of packets is (r, b)-constrained, with b > lmax . Find a bound on the end-to-end delay and delay variation. E XERCISE 2.2. Assume all nodes in a network are of the GR type with rate R and latency T . A flow with T-SPEC α(t) = min(rt + b, M + pt) has performed a reservation with rate R across a sequence of H nodes, with p ≥ R. Assume no reshaping is done. What is the buffer requirement at the hth node along the path, for h = 1, ...H ? E XERCISE 2.3. Assume all nodes in a network are made of a GR type with rate R and latency T , before which a re-shaper with shaping curve σ = γr,b is inserted. A flow with T-SPEC α(t) = min(rt + b, M + pt) has performed a reservation with rate R across a sequence of H such nodes, with p ≥ R. What is a buffer requirement at the hth node along the path, for h = 1, ...H ? E XERCISE 2.4. Assume all nodes in a network are made of a shaper followed by a FIFO multiplexer. Assume that flow I has T-SPEC, αi (t) = min(ri t + bi , M + pi t), that the shaper at every node uses the shaping curve σi = γri ,bi for flow i. Find the schedulability conditions for every node. E XERCISE 2.5. A network consists of two nodes in tandem. There are n1 flows of type 1 and n2 flows of type 2. Flows of type i have arrival curve αi (t) = ri t + bi , i = 1, 2. All flows go through nodes 1 then 2. Every node is made of a shaper followed by an EDF scheduler. At both nodes, the shaping curve for flows of type i is some σi and the delay budget for flows of type i is di . Every flow of type i should have a end-to-end delay bounded by Di . Our problem is to find good values of d1 and d2 . 1. We assume that σi = αi . What are the conditions on d1 and d2 for the end-to-end delay bounds to be satisfied ? What is the set of (n1 , n2 ) that are schedulable ? 2. Same question if we set σi = λri

98


E XERCISE 2.6. Consider the scheduler in Theorem 2.3.3. Find an efficient algorithm for computing the deadline of every packet. E XERCISE 2.7. Consider a SCED scheduler with target service curve for flow i given by βi = γri ,bi ⊗ δdi Find an efficient algorithm for computing the deadline of every packet. Hint: use an interpretation as a leaky bucket. E XERCISE 2.8. Consider the delay bound in Theorem 2.4.1. Take the same assumptions but assume also that the network is feedforward. Which better bound can you give ?

NETWORK CALCULUS Parts II and III A Theory of Deterministic Queuing Systems for the Internet J EAN -Y VES L E B OUDEC PATRICK T HIRAN

Online Version of the Book Springer Verlag - LNCS 2050 Reformatted for improved online viewing and printing Version May 10, 2004

99

100


PART II

M ATHEMATICAL BACKGROUND

101

C HAPTER 3

BASIC M IN - PLUS AND M AX - PLUS C ALCULUS In this chapter we introduce the basic results from Min-plus that are needed for the next chapters. Maxplus algebra is dual to Min-plus algebra, with similar concepts and results when minimum is replaced by maximum, and infimum by supremum. As basic results of network calculus use more min-plus algebra than max-plus algebra, we present here in detail the fundamentals of min-plus calculus. We briefly discuss the care that should be used when max and min operations are mixed at the end of the chapter. A detailed treatment of Min- and Max-plus algebra is provided in [28], here we focus on the basic results that are needed for the remaining of the book. Many of the results below can also be found in [11] for the discretetime setting.

3.1

M IN - PLUS C ALCULUS

In conventional algebra, the two most common operations on elements of Z or R are their addition and their multiplication. In fact, the set of integers or reals endowed with these two operations verify a number of well known axioms that define algebraic structures: (Z, +, ×) is a commutative ring, whereas (R, +, ×) is a field. Here we consider another algebra, where the operations are changed as follows: addition becomes computation of the minimum, multiplication becomes addition. We will see that this defines another algebraic structure, but let us first recall the notion of minimum and infimum.

3.1.1

I NFIMUM AND M INIMUM

Let S be a nonempty subset of R. S is bounded from below if there is a number M such that s ≥ M for all s ∈ S. The completeness axiom states that every nonempty subset S of R that is bounded from below has a greatest lower bound. We will call it infimum of S, and denote it by inf S. For example the closed and open intervals [a, b] and (a, b) have the same infimum, which is a. Now, if S contains an element that is smaller than all its other elements, this element is called minimum of S, and is denoted by min S. Note that the minimum of a set does not always exist. For example, (a, b) has no minimum since a∈ / (a, b). On the other hand, if the minimum of a set S exists, it is identical to its infimum. For example, min[a, b] = inf[a, b] = a. One easily shows that every finite nonempty subset of R has a minimum. Finally, let us mention that we will often use the notation ∧ to denote infimum (or, when it exists, the minimum). 103

CHAPTER 3. BASIC MIN-PLUS AND MAX-PLUS CALCULUS

104

For example, a ∧ b = min{a, b}. If S is empty, we adopt the convention that inf S = +∞. If f is a function from S to R, we denote by f (S) its range: f (S) = {t such that t = f (s) for some s ∈ S}. We will denote the infimum of this set by the two equivalent notations inf f (S) = inf {f (s)}. s∈S

We will also often use the following property. T HEOREM 3.1.1 (“F UBINI ” FORMULA FOR INFIMUM ). Let S be a nonempty subset of R, and f be a function from S to R. Let {Sn }n∈N be a collection of subsets of S, whose union is S. Then inf {f (s)} = inf inf {f (sn )} . s∈S

P ROOF :

s∈Sn

n∈N

By definition of an infimum, for any sets Sn , Sn = inf {inf Sn } . inf n

n

On the other hands, since ∪n Sn = S,

f

Sn

=

n∈N

f (Sn )

n∈N

so that

inf {f (s)} = inf f (S) = inf f

s∈S

= inf

f (Sn )

n∈N

=

3.1.2

inf

n∈N

Sn

n∈N

= inf {inf f (Sn )} n∈N

inf {f (s)} .

s∈Sn

D IOID (R ∪ {+∞}, ∧, +)

In traditional algebra, one is used to working with the algebraic structure (R, +, ×), that is, with the set of reals endowed with the two usual operations of addition and multiplication. These two operations possess a number of properties (associativity, commutativity, distributivity, etc) that make (R, +, ×) a commutative field. As mentioned above, in min-plus algebra, the operation of ‘addition’ becomes computation of the infimum (or of the minimum if it exists), whereas the one of ‘multiplication’ becomes the classical operation of addition. We will also include +∞ in the set of elements on which min-operations are carried out, so that the structure of interest is now (R ∪ {+∞}, ∧, +). Most axioms (but not all, as we will see later) defining a field still apply to this structure. For example, distribution of addition with respect to multiplication in conventional (‘Plus-times’) algebra (3 + 4) × 5 = (3 × 5) + (4 × 5) = 15 + 20 = 35

3.1. MIN-PLUS CALCULUS

105

translates in min-plus algebra as (3 ∧ 4) + 5 = (3 + 5) ∧ (4 + 5) = 8 ∧ 9 = 8. In fact, one easily verifies that ∧ and + satisfy the following properties: • (Closure of ∧) For all a, b ∈ R ∪ {+∞}, a ∧ b ∈ R ∪ {+∞}. • (Associativity of ∧) For all a, b, c ∈ R ∪ {+∞}, (a ∧ b) ∧ c = a ∧ (b ∧ c). • (Existence of a zero element for ∧) There is some e = +∞ ∈ R ∪ {+∞} such that for all a ∈ R ∪ {+∞}, a ∧ e = a. • (Idempotency of ∧) For all a ∈ R ∪ {+∞}, a ∧ a = a. • (Commutativity of ∧) For all a, b ∈ R ∪ {+∞}, a ∧ b = b ∧ a. • (Closure of +) For all a, b ∈ R ∪ {+∞}, a + b ∈ R ∪ {+∞}. • (Associativity of +) For all a, b, c ∈ R ∪ {+∞}, (a + b) + c = a + (b + c). • (The zero element for ∧ is absorbing for +) For all a ∈ R ∪ {+∞}, a + e = e = e + a. • (Existence of a neutral element for +) There is some u = 0 ∈ R ∪ {+∞} such that for all a ∈ R ∪ {+∞}, a + u = a = u + a. • (Distributivity of + with respect to ∧) For all a, b, c ∈ R ∪ {+∞}, (a ∧ b) + c = (a + c) ∧ (b + c) = c + (a ∧ b). A set endowed with operations satisfying all the above axioms is called a dioid. Moreover as + is also commutative (for all a, b ∈ R ∪ {+∞}, a + b = b + a), the structure (R ∪ {+∞}, ∧, +) is a commutative dioid. All the axioms defining a dioid are therefore the same axioms as the ones defining a ring, except one: the axiom of idempotency of the ‘addition’, which in dioids replaces the axiom of cancellation of ‘addition’ in rings (i.e. the existence of an element (−a) that ‘added’ to a gives the zero element). We will encounter other dioids later on in this chapter.

3.1.3

A C ATALOG OF W IDE - SENSE I NCREASING F UNCTIONS

A function f is wide-sense increasing if and only if f (s) ≤ f (t) for all s ≤ t. We will denote by G the set of non-negative wide-sense increasing sequences or functions and by F denote the set of wide-sense increasing sequences or functions such that f (t) = 0 for t < 0. Parameter t can be continuous or discrete: in the latter case, f = {f (t), t ∈ Z} is called a sequence rather than a function. In the former case, we take the convention that the function f = {f (t), t ∈ R} is left-continuous. The range of functions or sequences of F and G is R+ = [0, +∞]. Notation f + g (respectively f ∧ g) denotes the point-wise sum (resp. minimum) of functions f and g: (f + g)(t) = f (t) + g(t) (f ∧ g)(t) = f (t) ∧ g(t) Notation f ≤ (=, ≥)g means that f (t) ≤ (=, ≥)g(t) for all t. Some examples of functions belonging to F and of particular interest are the following ones. Notation [x]+ denotes max{x, 0}, x denotes the smallest integer larger than or equal to x. D EFINITION 3.1.1 (P EAK RATE FUNCTIONS λR ). Rt λR (t) = 0 for some R ≥ 0 (the ‘rate’).

if t > 0 otherwise


106

D EFINITION 3.1.2 (B URST DELAY FUNCTIONS δT ). δT (t) =

+∞ 0

if t > T otherwise

for some T ≥ 0 (the ‘delay’). D EFINITION 3.1.3 (R ATE - LATENCY FUNCTIONS βR,T ). βR,T (t) = R[t − T ] = +

R(t − T ) 0

if t > T otherwise

for some R ≥ 0 (the ‘rate’) and T ≥ 0 (the ‘delay’). D EFINITION 3.1.4 (A FFINE FUNCTIONS γr,b ). γr,b (t) =

rt + b 0

if t > 0 otherwise

for some r ≥ 0 (the ‘rate’) and b ≥ 0 (the ‘burst’). D EFINITION 3.1.5 (S TEP F UNCTION vT ). vT (t) = 1{t>T } =

1 0

if t > T otherwise

for some T > 0. D EFINITION 3.1.6 (S TAIRCASE F UNCTIONS uT,τ ). uT,τ (t) =

t+τ T

0

if t > 0 otherwise

for some T > 0 (the ‘interval’) and 0 ≤ τ ≤ T (the ‘tolerance’). These functions are also represented in Figure 3.1. By combining these basic functions, one obtains more general piecewise linear functions belonging to F. For example, the two functions represented in Figure 3.2 are written using ∧ and + from affine functions and rate-latency functions as follows, with r1 > r2 > . . . > rI and b1 < b2 < . . . < bI f1 = γr1 ,b1 ∧ γr2 ,b2 ∧ . . . γrI ,bI = min {γri ,bi } 1≤i≤I

(3.1)

f2 = λR ∧ {βR,2T + RT } ∧ {βR,4T + 2RT } ∧ . . . = inf {βR,2iT + iRT } . i≥0

(3.2)

We will encounter other functions later in the book, and obtain other representations with the min-plus convolution operator.


107

Burst-delay function

Peak rate function

δ (t) = 0 for t ≤ T T = ∞ for t > T

λ (t) = Rt R

R t

t T Affine function γ (t) = 0 for t = 0 r,b = rt + b for t > 0

Rate-latency function β

(t) = R[t-T]+

R,T

r

R b t

T

t

Staircase function v

Step function

(t) = ⎡(t+τ)/T⎤

u (t) = 1

T,τ

T

{t > T}

= 0 for t ≤ T 1 for t > T

4 3 2 1

1 T-τ

2T-τ

3T-τ

t

T

t

Figure 3.1: A catalog of functions of F : Peak rate function (top left), burst-delay function (top right), rate-latency function (center left), affine function (center right), staircase function (bottom left) and step function (bottom right).


108

f (t)

f (t)

2

1

r3 b3 b2 b1

r2 r1

2RT RT t

T 2T 3T

t

Figure 3.2: Two piecewise linear functions of F as defined by (3.1) (left) and (3.2) (right).

3.1.4

P SEUDO - INVERSE OF W IDE - SENSE I NCREASING F UNCTIONS

It is well known that any strictly increasing function is left-invertible. That is, if for any t1 < t2 , f (t1 ) < f (t2 ), then there is a function f −1 such that f −1 (f (t)) = t for all t. Here we consider slightly more general functions, namely, wide-sense increasing functions, and we will see that a pseudo-inverse function can defined as follows. D EFINITION 3.1.7 (P SEUDO - INVERSE ). Let f be a function or a sequence of F. The pseudo-inverse of f is the function (3.3) f −1 (x) = inf {t such that f (t) ≥ x} . For example, one can easily compute that the pseudo-inverses of the four functions of Definitions 3.1.1 to 3.1.4 are = λ1/R λ−1 R δT−1 = δ0 ∧ T

−1 βR,T

= γ1/R,T

−1 γr,b = β1/r,b .

The pseudo-inverse enjoys the following properties: T HEOREM 3.1.2 (P ROPERTIES OF PSEUDO - INVERSE FUNCTIONS ). Let f ∈ F, x, t ≥ 0. • (Closure) f −1 ∈ F and f −1 (0) = 0. • (Pseudo-inversion) We have that f (t) ≥ x ⇒ f −1 (x) ≤ t

(3.4)

(x) < t ⇒ f (t) ≥ x

(3.5)

f −1 (x) = sup {t such that f (t) < x} .

(3.6)

f • (Equivalent definition)

−1

Define subset Sx = {t such that f (t) ≥ x} ⊆ R+ . Then (3.3) becomes f −1 (x) = inf Sx . P ROOF : (Closure) Clearly, from (3.3), f −1 (x) = 0 for x ≤ 0 (and in particular f −1 (0) = 0). Now, let 0 ≤ x1 < x2 .


109

Then Sx1 ⊇ Sx2 , which implies that inf Sx1 ≤ inf Sx2 and hence that f −1 (x1 ) ≤ f −1 (x2 ). Therefore f −1 is wide-sense increasing. (Pseudo-inversion) Suppose first that f (t) ≥ x. Then t ∈ Sx , and so is larger than the infimum of Sx , which is f −1 (x): this proves (3.4). Suppose next that f −1 (x) < t. Then t > inf Sx , which implies that t ∈ Sx , by definition of an infimum. This in turn yields that f (t) ≥ x and proves (3.5). (Equivalent definition) Define subset S˜x = {t such that f (t) < x} ⊆ R+ . Pick t ∈ Sx and t˜ ∈ S˜x . Then f (t˜) < f (t), and since f is wide-sense increasing, it implies that t˜ ≤ t. This is true for any t ∈ Sx and t˜ ∈ S˜x , hence sup S˜x ≤ inf Sx . As S˜x ∪ Sx = R+ , we cannot have sup S˜x < inf Sx . Therefore sup S˜x = inf Sx = f −1 (x).

3.1.5

C ONCAVE , C ONVEX AND S TAR - SHAPED F UNCTIONS

As an important class of functions in min-plus calculus are the convex and concave functions, it is useful to recall some of their properties. D EFINITION 3.1.8 (C ONVEXITY IN Rn ). Let u be any real such that 0 ≤ u ≤ 1. • Subset S ⊆ Rn is convex if and only if ux + (1 − u)y ∈ S for all x, y ∈ S. • Function f from a subset D ⊆ Rn to R is convex if and only if f (ux+(1−u)y) ≤ uf (x)+(1−u)f (y) for all x, y ∈ D. • Function f from a subset D ⊆ Rn to R is concave if and only if −f is convex. For example, the rate-latency function (Fig 3.1, center left) is convex, the piecewise linear function f1 given by (3.1) is concave and the piecewise linear function f2 given by (3.2) is neither convex nor concave. There are a number of properties that convex sets and functions enjoy [76]. Here are a few that will be used in this chapter, and that are a direct consequence of Definition 3.1.8. • The convex subsets of R are the intervals. • If S1 and S2 are two convex subsets of Rn , their sum S = S1 + S2 = {s ∈ Rn | s = s1 + s2 for some s1 ∈ S1 and s2 ∈ S2 } is also convex. • Function f from an interval [a, b] to R is convex (resp. concave) if and only if f (ux + (1 − u)y) ≤ (resp. ≥) uf (x) + (1 − u)f (y) for all x, y ∈ [a, b] and all u ∈ [0.1]. • The pointwise maximum (resp. minimum) of any number of convex (resp. concave) functions is a convex (resp. concave) function. • If S is a convex subset of Rn+1 , n ≥ 1, the function from Rn to R defined by f (x) = inf{µ ∈ R such that (x, µ) ∈ S} is convex. • If f is a convex function from Rn to R, the set S defined by S = {(x, µ) ∈ Rn+1 such that f (x) ≤ µ} is convex. This set is called the epigraph of f . It implies in the particular case where n = 1 that the line segment between {a, f (a)} and {b, f (b)} lies above the graph of the curve y = f (x).

110


The proof of these properties is given in [76] and can be easily deduced from Definition 3.1.8, or even from a simple drawing. Chang [11] introduced star-shaped functions, which are defined as follows. D EFINITION 3.1.9 (S TAR - SHAPED FUNCTION ). Function f ∈ F is star-shaped if and only if f (t)/t is wide-sense decreasing for all t > 0. Star-shaped enjoy the following property: T HEOREM 3.1.3 (M INIMUM OF STAR - SHAPED FUNCTIONS ). Let f, g be two star-shaped functions. Then h = f ∧ g is also star-shaped. P ROOF : Consider some t ≥ 0. If h(t) = f (t), then for all s > t, h(t)/t = f (t)/t ≥ f (s)/s ≥ h(s)/s. The same argument holds of course if h(t) = g(t). Therefore h(t)/t ≥ h(s)/s for all s > t, which shows that h is star-shaped. We will see other properties of star-shaped functions in the next sections. Let us conclude this section with an important class of star-shaped functions. T HEOREM 3.1.4. Concave functions are star-shaped. Let f be a concave function. Then for any u ∈ [0, 1] and x, y ≥ 0, f (ux + (1 − u)y) ≥ P ROOF : uf (x) + (1 − u)f (y). Take x = t, y = 0 and u = s/t, with 0 < s ≤ t. Then the previous inequality becomes f (s) ≥ (s/t)f (t), which shows that f (t)/t is a decreasing function of t. On the other hand, a star-shaped function is not necessarily a concave function. We will see one such example in Section 3.1.7.

3.1.6

M IN - PLUS C ONVOLUTION

Let f (t) be a real-valued function, which is zero for t ≤ 0. If t ∈ R, the integral of this function in the conventional algebra (R, +, ×) is t f (s)ds 0

which becomes, for a sequence f (t) where t ∈ Z, t

f (s).

s=0

In the min-plus algebra (R ∪ {+∞}, ∧, +), where the ‘addition’ is ∧ and the ‘multiplication’ is +, an ‘integral’ of the function f becomes therefore {f (s)},

inf


which becomes, for a sequence f (t) where t ∈ Z, min

{f (s)}.

s∈Z such that 0≤s≤t

We will often adopt a shorter notation for the two previous expressions, which is inf {f (s)},

0≤s≤t

with s ∈ Z or s ∈ R depending on the domain of f .


111

A key operation in conventional linear system theory is the convolution between two functions, which is defined as +∞ (f ⊗ g)(t) = f (t − s)g(s)ds −∞

and becomes, when f (t) and g(t) are two functions that are zero for t < 0, t f (t − s)g(s)ds. (f ⊗ g)(t) = 0

In min-plus calculus, the operation of convolution is the natural extension of the previous definition: D EFINITION 3.1.10 (M IN - PLUS CONVOLUTION ). Let f and g be two functions or sequences of F. The min-plus convolution of f and g is the function (f ⊗ g)(t) = inf {f (t − s) + g(s)} . 0≤s≤t

(3.7)

(If t < 0, (f ⊗ g)(t) = 0). Example. Consider the two functions γr,b and βR,T , with 0 < r < R, and let us compute their min-plus convolution. Let us first compute it for 0 ≤ t ≤ T . (γr,b ⊗ βR,T )(t) = inf γr,b (t − s) + R[s − T ]+ 0≤s≤t

=

inf {γr,b (t − s) + 0} = γr,b (0) + 0 = 0 + 0 = 0

0≤s≤t

Now, if t > T , one has (γr,b ⊗ βR,T )(t) = inf γr,b (t − s) + R[s − T ]+ 0≤s≤t = inf γr,b (t − s) + R[s − T ]+ ∧ inf γr,b (t − s) + R[s − T ]+ 0≤s≤T T ≤s RT and t ≤ nT , βR,nT + nK = nK > (n − 1)RT + K = R(nT − T ) + K ≥ R[t − T ]+ + K = βR,T + K ,


119

whereas if t > nT βR,nT + nK = R(t − nT ) + nK = R(t − T ) + (n − 1)(K − RT ) + K > R(t − T ) + K = βR,T + K so that (βR,T + K )(n) ≥ βR,T + K for all n ≥ 1. Therefore (3.13) becomes

βR,T + K = δ0 ∧ inf (βR,T + K )(n) = δ0 ∧ (βR,T + K ), n≥1

and is shown on the left of Figure 3.6. On the other hand, if K = K < RT , the infimum in the previous equation is not reached in n = 1 for every t > 0, so that the sub-additive closure is now expressed by

βR,T + K = δ0 ∧ inf (βR,T + K )(n) = δ0 ∧ inf (βR,nT + nK ) , n≥1

n≥1

and is shown on the right of Figure 3.6.

βR,T(t) + K’

βR,T(t) + K”

4K” 3K” 2K” K”

K’ RT t

T

T 2T 3T 4T

t

Figure 3.6: The sub-additive closure of functions βR,T + K (left) and βR,T + K (right), when K < RT < K . Among all the sub-additive functions that are smaller than f and that are zero in t = 0, there is one that is an upper bound for all others; it is equal to f , as established by the following theorem. T HEOREM 3.1.10 (S UB - ADDITIVE CLOSURE ). Let f be a function or a sequence of F, and let f be its sub-additive closure. Then (i) f ≤ f , f ∈ F and f is sub-additive. (ii) if function g ∈ F is sub-additive, with g(0) = 0 and g ≤ f , then g ≤ f . P ROOF : (i) It is obvious from Definition 3.1.12, that f ≤ f . By repeating (n − 1) times Rule 1 of Theorem 3.1.5, one has that f (n) ∈ F for all n ≥ 1. As f (0) = δ0 ∈ F too, f = inf n≥0 {f (n) } ∈ F. Let us show next that f is sub-additive. For any integers n, m ≥ 0, and for any s, t ≥ 0, f (n+m) (t + s) = (f (n) ⊗ f (m) )(t + s) =

inf

{f (n) (t + s − u) + f (m) (u)}

0≤u≤t+s

≤ f (n) (t) + f (m) (s) so that f (t + s) = ≤ =

inf {f (n+m) (t + s)} = inf {f (n+m) (t + s)}

n+m≥0

inf {f

n,m≥0

n,m≥0

(n)

(t) + f

(m)

(s)}

inf {f (n) (t)} + inf {f (m) (s)} = f (t) + f (s)

n≥0

m≥0


120

which shows that f is sub-additive. (ii) Next, suppose that g ∈ F is sub-additive, g(0) = 0 and g ≤ f . Suppose that for some n ≥ 1, f (n) ≥ g. Clearly, this holds for n = 0 (because g(0) = 0 implies that g ≤ δ0 = f (0) ) and for n = 1. Now, this assumption and the sub-additivity of g yield that for any 0 ≤ s ≤ t, f (n) (t − s) + f (s) ≥ g(t − s) + g(s) ≥ g(t) and hence that f (n+1) (t) ≥ g(t). By recursion on n, f (n) ≥ g for all n ≥ 0, and therefore f = inf n≥0 {f (n) } ≥ g. C OROLLARY 3.1.1 (S UB - ADDITIVE CLOSURE OF A SUB - ADDITIVE FUNCTION ). Let f ∈ F. Then the three following statements are equivalent: (i) f (0) = 0 and f is sub-additive (ii) f ⊗ f = f (iii) f = f . P ROOF : (i) ⇒ (ii) follows immediately from from Definition 3.1.11. (ii) ⇒ (iii): first note that f ⊗f = f implies that f (n) = f for all n ≥ 1. Second, note that (f ⊗ f )(0) = f (0) + f (0), which implies that f (0) = 0. Therefore f = inf n≥0 {f (n) } = δ0 ∧ f = f . (iii) ⇒ (i) follows from Theorem 3.1.10. The following theorem establishes some additional useful properties of the sub-additive closure of a function. T HEOREM 3.1.11 (OTHER PROPERTIES OF SUB - ADDITIVE CLOSURE ). Let f, g ∈ F • (Isotonicity) If f ≤ g then f ≤ g. • (Sub-additive closure of a minimum) f ∧ g = f ⊗ g. • (Sub-additive closure of a convolution) f ⊗ g ≥ f ⊗ g. If f (0) = g(0) = 0 then f ⊗ g = f ⊗ g. P ROOF : (Isotonocity) Suppose that we have shown that for some n ≥ 1, f (n) ≥ g (n) (Clearly, this holds for n = 0 and for n = 1). Then applying Theorem 3.1.7 we get f (n+1) = f (n) ⊗ f ≥ g (n) ⊗ g = g (n+1) , which implies by recursion on n that f ≤ g. (Sub-additive closure of a minimum) One easily shows, using Theorem 3.1.5, that (f ∧ g)(2) = (f ⊗ f ) ∧ (f ⊗ g) ∧ (g ⊗ g). Suppose that we have shown that for some n ≥ 0, the expansion of (f ∧ g)(n) is (f ∧ g)(n) = f (n) ∧ (f (n−1) ⊗ g) ∧ (f (n−2) ⊗ g (2) ) ∧ . . . ∧ g (n) =

f (n−k) ⊗ g (k) . inf 0≤k≤n

Then

(f ∧ g)(n+1) = (f ∧ g) ⊗ (f ∧ g)(n) = f ⊗ (f ∧ g)(n) ∧ g ⊗ (f ∧ g)(n)

= inf f (n+1−k) ⊗ g (k) ∧ inf f (n−k) ⊗ g (k+1) 0≤k≤n 0≤k≤n

(n+1−k) (k) (n+1−k ) (k ) f ∧ inf f ⊗g ⊗ g = inf 0≤k≤n 1≤k ≤n+1

f (n+1−k) ⊗ g (k) = inf 0≤k≤n+1

which establishes the recursion for all n ≥ 0. Therefore

f ∧ g = inf inf f (n−k) ⊗ g (k) = inf inf f (n−k) ⊗ g (k) n≥0 0≤k≤n k≥0 n≥k

(l) (k) (l) (k) = inf inf {f } ⊗ g = inf inf f ⊗ g k≥0 l≥0 k≥0 l≥0

= inf f ⊗ g (k) = f ⊗ inf {g (k) } = f ⊗ g. k≥0

k≥0


121

(Sub-additive closure of a convolution) Using the same recurrence argument as above, one easily shows that (f ⊗ g)(n) = f (n) ⊗ g (n) , and hence that

f ⊗ g = inf (f ⊗ g)(n) = inf f (n) ⊗ g (n) n≥0 n≥0

(n) (m) ≥ inf f ⊗ g n,m≥0

(n) (m) ⊗ inf g = f ⊗ g. (3.14) = inf f n≥0

m≥0

If f (0) = g(0) = 0, Rule 8 in Theorem 3.1.6 yields that f ⊗ g ≤ f ∧ g, and therefore that f ⊗ g ≤ f ∧ g. Now we have just shown above that f ∧ g = f ⊗ g, so that f ⊗ g ≤ f ⊗ g. Combining this result with (3.14), we get f ⊗ g = f ⊗ g. Let us conclude this section with an example illustrating the effect that a difference in taking t continuous or discrete may have. This example is the computation of the sub-additive closure of 2 if t > 0 t f (t) = 0 if t ≤ 0 Suppose first that t ∈ R. Then we compute that (f ⊗ f )(t) = inf (t − s)2 + s2 = (t/2)2 + (t/2)2 = t2 /2 0≤s≤t

as the infimum is reached in s = t/2. By repeating this operation n times, we obtain

f (n) (t) = inf (t − s)2 + (f (n−1) )2 (s) = 0≤s≤t inf (t − s)2 + s2 /(n − 1) = t2 /n 0≤s≤t

as the infimum is reached in s = t(1 − 1/n). Therefore f (t) = inf {t2 /n} = lim t2 /n = 0. n→∞

n≥0

Consequently, if t ∈ R, the sub-additive closure of function f is f = 0, as shown on the left of Figure 3.7. Now, if t ∈ Z, the sequence f (t) is convex and piecewise linear, as we can always connect the different successive points (t, t2 ) for all t = 0, 1, 2, 3, . . .: the resulting graph appears as a succession of segments of slopes equal to (2t + 1) (the first segment in particular has slope 1), and of projections on the horizontal axis having a length equal to 1, as shown on the right of Figure 3.7. Therefore we can apply Rule 9 of Theorem 3.1.6, which yields that f ⊗ f is obtained by doubling the length of the different linear segments of f , and putting them end-to-end by increasing slopes. The analytical expression of the resulting sequence is (f ⊗ f )(t) = min (t − s)2 + s2 = t2 /2 . 0≤s≤t

Sequence f (2) = f ⊗ f is again convex and piecewise linear. Note the first segment has slope 1, but has now a double length. If we repeat n times this convolution, it will result in a convex, piecewise linear sequence f (n) (t) whose first segment has slope 1 and horizontal length n: f (n) (t) = t

if 0 ≤ t ≤ n,


122

f(t)

f(t) f

(2)

f

(t) f

(n)

(t)

f

(n)

(t)

f(t)

(t)

f(t)

(2)

1

t

12 3

t

Figure 3.7: The sub-additive closure of f (t) = tλ1 (t), when t ∈ R (left) and when t ∈ Z (right). as shown on the right of Figure 3.7. Consequently, the sub-additive closure of sequence f is obtained by letting n → ∞, and is therefore f (t) = t for t ≥ 0. Therefore, if t ∈ Z, f = λ1 .

3.1.9

M IN - PLUS D ECONVOLUTION

The dual operation (in a sense that will clarified later on) of the min-plus convolution is the min-plus deconvolution. Similar considerations as the ones of Subsection 3.1.1 can be made on the difference between a sup and a max. Notation ∨ stands for sup or, if it exists, for max: a ∨ b = max{a, b}. D EFINITION 3.1.13 (M IN - PLUS DECONVOLUTION ). Let f and g be two functions or sequences of F. The min-plus deconvolution of f by g is the function (f g)(t) = sup {f (t + u) − g(u)} .

(3.15)

u≥0

If both f (t) and g(t) are infinite for some t, then Equation (3.15) is not defined. Contrary to min-plus convolution, function (f g)(t) is not necessarily zero for t ≤ 0, and hence this operation is not closed in F, as shown by the following example. Example. Consider again the two functions γr,b and βR,T , with 0 < r < R, and let us compute the min-plus deconvolution of γr,b by βR,T . We have that (γr,b βR,T )(t) = sup γr,b (t + u) − R[u − T ]+ u≥0 = sup γr,b (t + u) − R[u − T ]+ ∨ sup γr,b (t + u) − R[u − T ]+ 0≤u≤T

=

u>T

sup {γr,b (t + u)} ∨ sup {γr,b (t + u) − Ru + RT }

0≤u≤T

u>T

= {γr,b (t + T )} ∨ sup {γr,b (t + u) − Ru + RT } . u>T

Let us first compute this expression for t ≤ −T . Then γr,b (t + T ) = 0 and (3.16) becomes (γr,b βR,T )(t)

(3.16)

3.1. MIN-PLUS CALCULUS = 0∨

123

sup {γr,b (t + u) − Ru + RT }

T −t

= 0∨

sup {0 − Ru + RT } ∨ sup {b + r(t + u) − Ru + RT } u>−t

T −T . Then (3.16) becomes (γr,b βR,T )(t) = {b + r(t + T )} ∨ sup {b + r(t + u) − Ru + RT } u>T

= {b + r(t + T )} ∨ {b + r(t + T )} = b + r(t + T ). The result is shown in Figure 3.8.

(γr,b ∅ βR,T )(t) r b R t

–T Figure 3.8: Function γr,b βR,T when 0 < r < R.

Let us now state some properties of (Other properties will be given in the next section). T HEOREM 3.1.12 (P ROPERTIES OF ). Let f, g, h ∈ F. • • • • •

Rule 11 (Isotonicity of ) If f ≤ g, then f h ≤ g h and h f ≥ h g. Rule 12 (Composition of ) (f g) h = f (g ⊗ h). Rule 13 (Composition of and ⊗) (f ⊗ g) g ≤ f ⊗ (g g). Rule 14 (Duality between and ⊗) f g ≤ h if and only if f ≤ g ⊗ h. Rule 15 (Self-deconvolution) (f f ) is a sub-additive function of F such that (f f )(0) = 0.

P ROOF :

(Rule 11) If f ≤ g, then for any h ∈ F (f h)(t) = sup {f (t + u) − h(u)} ≤ sup {g(t + u) − h(u)} = (g h)(t) u≥0

u≥0

(h f )(t) = sup {h(t + u) − f (u)} ≥ sup {h(t + u) − g(u)} = (h g)(t). u≥0

(Rule 12) One computes that

u≥0


124

((f g) h)(t) = sup {(f g)(t + u) − h(u)} u≥0 = sup sup {f (t + u + v) − g(v)} − h(u) u≥0 v≥0 = sup sup f (t + v ) − g(v − u) − h(u) v ≥u

u≥0

= sup sup f (t + v ) − g(v − u) + h(u) u≥0 v ≥u

= sup sup

v ≥0 0≤u≤v

f (t + v ) − g(v − u) + h(u)

= sup f (t + v ) − inf g(v − u) + h(u) v ≥0

0≤u≤v

= sup f (t + v ) − (g ⊗ h)(v ) = (f (g ⊗ h))(t).

v ≥0

(Rule 13) One computes that ((f ⊗ g) g)(t) = sup {(f ⊗ g)(t + u) − g(u)} u≥0

{f (t + u − s) + g(s) − g(u)} f (t − s ) + g(s + u) − g(u) sup inf u≥0 −u≤s ≤t sup inf f (t − s ) + g(s + u) − g(u) u≥0 0≤s ≤t f (t − s ) + sup{g(s + v) − g(v)} sup inf u≥0 0≤s ≤t v≥0 f (t − s ) + sup{g(s + v) − g(v)} inf 0≤s ≤t v≥0 f (t − s ) + (g g)(s ) = (f ⊗ (g g))(t). inf

= sup

inf

u≥0 0≤s≤t+u

= ≤ ≤ = =

0≤s ≤t

(Rule 14) Suppose first that (f g)(s) ≤ h(s) for all s. Take any s, v ≥ 0. Then f (s + v) − g(v) ≤ sup {f (s + u) − g(u)} = (f g)(s) ≤ h(s) u≥0

or equivalently, f (s + v) ≤ g(v) + h(s). Let t = s + v. The former inequality can be written as f (t) ≤ g(t − s) + h(s). As it is verified for all t ≥ s ≥ 0, it is also verified in particular for the value of s that achieves the infimum of the right-hand side of this inequality. Therefore it is equivalent to f (t) ≤ inf {g(t − s) + h(s)} = (g ⊗ h)(t) 0≤s≤t

for all t ≥ 0. Suppose now that for all v, f (v) ≤ (g ⊗ h)(v). Pick any t ∈ R. Then, since g, h ∈ F, f (v) ≤ inf {g(v − s) + h(s)} = inf {g(v − s) + h(s)} ≤ g(t − v) + h(t). 0≤s≤v

s∈R


125

Let u = t − v, the former inequality can be written as f (t + u) − g(u) ≤ h(t). As this is true for all u, it is also verified in particular for the value of u that achieves the supremum of the left-hand side of this inequality. Therefore it is equivalent to sup {f (t + u) − g(u)} ≤ h(t). u∈R

Now if u < 0, g(u) = 0, so that supu T such that f (t) ≥ L for all t > T1 . Now let t > 2T1 . If u > T1 , then f (u) ≥ L. Otherwise, u ≤ T1 thus t − u ≥ t − T1 > T1 thus g(t − u) ≥ L. Thus in all cases f (u) + g(t − u) ≥ L. Thus we have shown that (g ⊗ f )(t) ≥ L for t > 2T1 .

(3.18)

Combining (3.17) and (3.18) shows the lemma. D EFINITION 3.1.14 (T IME I NVERSION ). For a fixed T ∈ [0, +∞[, the inversion operator ΦT is defined on G by: ΦT (f )(g) = g(+∞) − g(T − t) Graphically, time inversion can be obtained by a rotation of 180o around the point ( T2 , g(+∞) ). It is simple 2 that time inversion is symmetrical (ΦT (ΦT (g)) = g) and preserves the total to check that ΦT (g) is in G, value (ΦT (g)(+∞) = g(+∞)). Lastly, for any α and T , α is an arrival curve for g if and only if α is an arrival curve for ΦT (g). and let T T HEOREM 3.1.14 (R EPRESENTATION OF D ECONVOLUTION BY T IME I NVERSION ). Let g ∈ G, be such that g(T ) = g(+∞). Let f ∈ F be such that limt→+∞ f (t) = +∞. Then g f = ΦT (ΦT (g) ⊗ f )

(3.19)

The theorem says that g f can be computed by first inverting time, then computing the min-plus convolution between f , and the time-inverted function g, and then inverting time again. Figure 3.9 shows a graphical illustration. P ROOF : The proof consists in computing the right handside in Equation (3.19). Call gˆ = ΦT (g). We have, by definition of the inversion ΦT (ΦT (g) ⊗ f ) = ΦT (ˆ g ⊗ f ) = (ˆ g ⊗ f )(+∞) − (ˆ g ⊗ f )(T − t) Now from Lemma 3.1.1 and the preservation of total value: (ˆ g ⊗ f )(+∞) = gˆ(+∞) = g(+∞) Thus, the right-handside in Equation (3.19) is equal to g(+∞) − (ˆ g ⊗ f )(T − t) = g(+∞) − inf {ˆ g (T − t − u) + f (u)} u≥0

Again by definition of the inversion, it is equal to g(+∞) − inf {g(+∞) − g(t + u) + f (u)} = sup{g(t + u) − f (u)}. u≥0

u≥0


127

g(t) f(t)

t T

T

0

g(t) g(T)/2

ΦT (g)(t) t T/2

T

ΦT (g)(t) (ΦT (g) ⊗ f )(t) t

(ΦT (ΦT (g) ⊗ f ))(t) = (g ∅

f)(t)

(ΦT (g) ⊗ f )(t)

g(T)/2

t T/2

T

Figure 3.9: Representation of the min-plus deconvolution of g by f = γr,b by time-inversion. From top to bottom: functions f and g, function ΦT (g), function ΦT (g) ⊗ f and finally function g f = ΦT (ΦT (g) ⊗ f ).


128

3.1.11

V ERTICAL AND H ORIZONTAL D EVIATIONS

The deconvolution operator allows to easily express two very important quantities in network calculus, which are the maximal vertical and horizontal deviations between the graphs of two curves f and g of F. The mathematical definition of these two quantities is as follows. D EFINITION 3.1.15 (V ERTICAL AND HORIZONTAL DEVIATIONS ). Let f and g be two functions or sequences of F. The vertical deviation v(f, g) and horizontal deviation h(f, g) are defined as v(f, g) = sup {f (t) − g(t)}

(3.20)

h(f, g) = sup {inf {d ≥ 0 such that f (t) ≤ g(t + d)}} .

(3.21)

t≥0 t≥0

Figure 3.10 illustrates these two quantities on an example.

g(t) h(f,g) f(t) v(f,g) t Figure 3.10: The horizontal and vertical deviations between functions f and g. Note that (3.20) can be recast as v(f, g) = (f g)(0)

(3.22)

whereas (3.20) is equivalent to requiring that h(f, g) is the smallest d ≥ 0 such that for all t ≥ 0, f (t) ≤ g(t + d) and can therefore be recast as h(f, g) = inf {d ≥ 0 such that (f g)(−d) ≤ 0} . Now the horizontal deviation can be more easily computed from the pseudo-inverse of g. Indeed, Definition 3.1.7 yields that g −1 (f (t)) = inf {∆ such that g(∆) ≥ f (t)} = inf {d ≥ 0 such that g(t + d) ≥ f (t)} + t so that (3.21) can be expressed as

h(f, g) = sup g −1 (f (t)) − t = (g −1 (f ) λ1 )(0). t≥0

We have therefore the following expression of the horizontal deviation between f and g: P ROPOSITION 3.1.1 (H ORIZONTAL DEVIATION ).

h(f, g) = sup g −1 (f (t)) − t . t≥0

(3.23)

3.2. MAX-PLUS CALCULUS

3.2

129

M AX - PLUS C ALCULUS

Similar definitions, leading to similar properties, can be derived if we replace the infimum (or minimum, it is exists) by a supremum (or maximum, if it exists). We use the notation ∨ for denoting sup or max. In particular, one can show that (R ∪ {−∞}, ∨, +) is also a dioid, and construct a max-plus convolution and deconvolution, which are defined as follows.

3.2.1

M AX - PLUS C ONVOLUTION AND D ECONVOLUTION

D EFINITION 3.2.1 (M AX - PLUS CONVOLUTION ). Let f and g be two functions or sequences of F. The max-plus convolution of f and g is the function (f ⊗g)(t) = sup {f (t − s) + g(s)} .

(3.24)

0≤s≤t

(If t < 0, (f ⊗g)(t) = 0). D EFINITION 3.2.2 (M AX - PLUS DECONVOLUTION ). Let f and g be two functions or sequences of F. The max-plus deconvolution of f by g is the function (f g)(t) = inf {f (t + u) − g(u)} . u≥0

3.2.2

(3.25)

L INEARITY OF M IN - PLUS D ECONVOLUTION IN M AX - PLUS A LGEBRA

Min-plus deconvolution is, in fact, an operation that is linear in (R+ , ∨, +). Indeed, one easily shows the following property. T HEOREM 3.2.1 (L INEARITY OF IN MAX - PLUS ALGEBRA ). Let f, g, h ∈ F. • Rule 16 (Distributivity of with respect to ∨) (f ∨ g) h = (f h) ∨ (g h). • Rule 17 (Addition of a constant) For any K ∈ R+ , (f + K) g = (f g) + K. Min-plus convolution is not, however, a linear operation in (R+ , ∨, +), because in general (f ∨ g) ⊗ h = (f ⊗ h) ∨ (g ⊗ h). Indeed, take f = β3R,T , g = λR and h = λ2R for some R, T > 0. Then using Rule 9, one easily computes (see Figure 3.11) that f ⊗ h = β3R,T ⊗ λ2R = β2R,T g ⊗ h = λR ⊗ λ2R = λR (f ∨ g) ⊗ h = (β3R,T ∨ λR ) ⊗ λ2R = β2R,3T /4 ∨ λR = β2R,T ∨ λR = (f ⊗ h) ∨ (g ⊗ h). Conversely, we have seen that min-plus convolution is a linear operation in (R+ , ∧, +), and one easily shows that min–plus deconvolution is not linear in (R+ , ∧, +). Finally, let us mention that one can also replace + by ∧, and show that (R ∪ {+∞} ∪ {−∞}, ∨, ∧) is also a dioid. Remark However, as we have seen above, as soon as the three operations ∧, ∨ and + are involved in a computation, one must be careful before applying any distribution.


130

((f⊗h)∨(g⊗h))(t)

((f∨g)⊗h))(t)

2R

2R 2RT 3RT/2 R R T

2T

t

3T/4 3T/2

t

Figure 3.11: Function (f ⊗ h) ∨ (g ⊗ h) (left) and (f ∨ g) ⊗ h (right) when f = β3R,T , g = λR and h = λ2R for some R, T > 0.

3.3

E XERCISES

E XERCISE 3.1. 1. Compute α ⊗ δ for any function α 2. Express the rate-latency function by means of δ and λ functions. 1. Compute i βi when βi is a rate-latency function E XERCISE 3.2. 2. Compute β1 ⊗ β2 with β1 (t) = R(t − T )+ and β2 (t) = (rt + b)1{t>0} E XERCISE 3.3.

1. Is ⊗ distributive with respect to the min operator ?

C HAPTER 4

M IN - PLUS AND M AX -P LUS S YSTEM T HEORY In Chapter 3 we have introduced the basic operations to manipulate functions and sequences in Min-Plus or Max-Plus algebra. We have studied in detail the operations of convolution, deconvolution and sub-additive closure. These notions form the mathematical cornerstone on which a first course of network calculus has to be built. In this chapter, we move one step further, and introduce the theoretical tools to solve more advanced problems in network calculus developed in the second half of the book. The core object in Chapter 3 were functions and sequences on which operations could be performed. We will now place ourselves at the level of operators mapping an input function (or sequence) to an output function or sequence. Max-plus system theory is developed in detail in [28], here we focus on the results that are needed for the remaining chapters of the book. As in Chapter 3, we focus here Min-Plus System Theory, as Max-Plus System Theory follows easily by replacing minimum by maximum, and infimum by supremum.

4.1 4.1.1

M IN -P LUS AND M AX -P LUS O PERATORS V ECTOR N OTATIONS

Up to now, we have only worked with scalar operations on scalar functions in F or G. In this chapter, we will also work with vectors and matrices. The operations are extended in a straightforward manner. Let J be a finite, positive integer. For vectors z, z ∈ R+ J , we define z ∧z as the coordinate-wise minimum of z and z , and similarly for the + operator. We write z ≤ z with the meaning that zj ≤ zj for 1 ≤ j ≤ J. Note that the comparison so defined is not a total order, that is, we cannot guarantee that either z ≤ z or z ≤ z holds. For a constant K, we note z + K the vector defined by adding K to all elements of z. We denote by G J the set of J-dimensional wide-sense increasing real-valued functions or sequences of parameter t, and F J the subset of functions that are zero for t < 0. For sequences or functions x(t), we note similarly (x ∧ y )(t) = x(t) ∧ y (t) and (x + K)(t) = x(t) + K for all t ≥ 0, and write x ≤ y with the meaning that x(t) ≤ y (t) for all t. For matrices A, B ∈ R+ J × R+ J , we define A ∧ B as the entry-wise minimum of A and B. For vector z ∈ R+ J , the ‘multiplication’ of vector z ∈ R+ J by matrix A is – remember that in min-plus algebra, 131

CHAPTER 4. MIN-PLUS AND MAX-PLUS SYSTEM THEORY

132 multiplication is the + operation – by

A + z, and has entries min1≤j≤J (aij + zj ). Likewise, the ‘product’ of two matrices A and B is denoted by A + B and has entries min1≤j≤J (aij + bjk ) for 1 ≤ i, k ≤ J. Here is an example of a ‘multiplication’ of a vector by a matrix, when J = 2 5 3 2 4 + = 1 3 1 3 and an example of a matrix ‘multiplication’ is 5 3 2 4 4 3 + = . 1 3 1 0 3 3 2

We denote by F J the set of J × J matrices whose entries are functions or sequences of F, and similarly 2 for G J . 2

The min-plus convolution of a matrix A ∈ F J by a vector z ∈ F J is the vector of F J defined by (A ⊗ z)(t) = inf {A(t − s) + z(s)} 0≤s≤t

and whose J coordinates are thus min {aij ⊗ zj }(t) = inf

min {aij (t − s) + zj (s)}.

0≤s≤t 1≤j≤J

1≤j≤J

Likewise, A ⊗ B is defined by (A ⊗ B)(t) = inf {A(t − s) + B(s)} 0≤s≤t

and has entries min1≤j≤J (aij ⊗ bjk ) for 1 ≤ i, k ≤ J. For example, we have

and

λr ∞ ∞ δT

λr ∞ ∞ δT

⊗

⊗

γr/2,b δ2T

γr/2,b γr,b δ2T λr

=

=

λr ∧ γr/2,b δ3T

λr ∧ γr/2,b λr δ3T βr,T

.

Finally, we will also need to extend the set of wide-sense increasing functions G to include non decreasing functions of two arguments. We adopt the following definition (a slightly different definition can be found in [11]). D EFINITION 4.1.1 (B IVARIATE WIDE - SENSE INCREASING FUNCTIONS ). We denote by G˜ the set of bivariate functions (or sequences) such that for all s ≤ s and any t ≤ t f (t, s) ≤ f (t, s ) f (t, s) ≤ f (t , s). We call such functions bivariate wide-sense increasing functions. In the multi-dimensional case, we denote by G˜J the set of J × J matrices whose entries are wide-sense 2 increasing bivariate functions. A matrix of A(t) ∈ F J is a particular case of a matrix H(t, s) ∈ G˜J , with s set to a fixed value.

4.1. MIN-PLUS AND MAX-PLUS OPERATORS

4.1.2

133

O PERATORS

A system is an operator Π mapping an input function or sequence x onto an output function or sequence y = Π(x). We will always assume in this book that x, y ∈ G J , where J is a fixed, finite, positive integer. This means that each of the J coordinates xj (t), yj (t), 1 ≤ j ≤ J, is a wide-sense increasing function (or sequence) of t. It is important to mention that Min-plus system theory applies to more general operators, taking RJ to RJ , where neither the input nor the output functions are required to be wide-sense increasing. This requires minor modifications in the definitions and properties established in this chapter, see [28] for the theory described in a more general setting. In this book, to avoid the unnecessary overhead of new notations and definitions, we decided to expose min-plus system theory for operators taking G J to G J . Most often, the only operator whose output may not be in F J is deconvolution, but all other operators we need will take F J to F J . Most of the time, the dimension of the input and output is J = 1, and the operator takes F to F. We will speak of a scalar operator. In this case, we will drop the arrow on the input and output, and write y = Π(x) instead. We write Π1 ≤ Π2 with the meaning that Π1 (x) ≤ Π2 (x) for all x, which in turn has the meaning that Π1 (x)(t) ≤ Π2 (x)(t) for all t. For a set of operators Πs , indexed by s in some set S, we call inf s∈S Πs the operator defined by [inf s∈S Πs ](x(t)) = inf s∈S [Πs (x(t))]. For S = {1, 2} we denote it with Π1 ∧ Π2 . We also denote by ◦ the composition of two operators: (Π1 ◦ Π2 )(x) = Π1 (Π2 (x)). We leave it to the alert reader to check that inf s∈S Πs and Π1 ◦ Π2 do map functions in G J to functions in GJ .

4.1.3

A C ATALOG OF O PERATORS

Let us mention a few examples of scalar operators of particular interest. The first two have already been studied in detail in Chapter 3, whereas the third was introduced in Section 1.7. The fact that these operators map G J into G J follows from Chapter 3. D EFINITION 4.1.2 (M IN - PLUS CONVOLUTION Cσ ). Cσ :

F → F x(t) → y(t) = Cσ (x)(t) = (σ ⊗ x)(t) = inf 0≤s≤t {σ(t − s) + x(s)} ,

for some σ ∈ F. D EFINITION 4.1.3 (M IN - PLUS DECONVOLUTION Dσ ). Dσ :

F → G x(t) → y(t) = Dσ (x)(t) = (x σ)(t) = supu≥0 {x(t + u) − σ(u)} ,

for some σ ∈ F. Note that Min-plus deconvolution produces an output that does not always belong to F.


134

D EFINITION 4.1.4 (PACKETIZATION PL ). PL :

F → F x(t) → y(t) = PL (x)(t) = P L (x(t)) = supi∈N L(i)1L(i)≤x(t) ,

for some wide-sense increasing sequence L (defined by Definition 1.7.1). We will also need later on the following operator, whose name will be justified later in this chapter. D EFINITION 4.1.5 (L INEAR IDEMPOTENT OPERATOR hσ ). hσ :

F → F x(t) → y(t) = hσ (x)(t) = inf 0≤s≤t {σ(t) − σ(s) + x(s)} ,

for some σ ∈ F. The extension of the scalar operators to the vector case is straightforward. The vector extension of the convolution is for instance: D EFINITION 4.1.6 (V ECTOR MIN - PLUS CONVOLUTION CΣ ). CΣ :

FJ → FJ x(t) → y (t) = CΣ (x)(t) = (Σ ⊗ x)(t) = inf 0≤s≤t {Σ(t − s) + x(s)} ,

2

for some Σ ∈ F J . If the (i, j)th entry of Σ is σij , the ith component of y (t) reads therefore yi (t) = inf

min {σij (t − s) + xj (s)}

0≤s≤t 1≤j≤J

Let us conclude with the shift operator, which we directly introduce in the vector setting: D EFINITION 4.1.7 (S HIFT OPERATOR ST ). ST

:

GJ → GJ x(t) → y (t) = ST (x)(t) = x(t − T ),

for some T ∈ R. Let us remark that S0 is the identity operator: S0 (x) = x.

4.1.4

U PPER AND L OWER S EMI -C ONTINUOUS O PERATORS

We now study a number of properties of min-plus linear operators. We begin with that of upper-semi continuity. D EFINITION 4.1.8 (U PPER SEMI - CONTINUOUS OPERATOR ). Operator Π is upper semi-continuous if for any (finite or infinite) set of functions or sequences {xn }, xn ∈ G J , Π inf {xn } = inf {Π(xn )} . n

n

(4.1)


135

We can check that Cσ , CΣ , hσ and ST are upper semi-continuous. For example, for CΣ , we check indeed that

CΣ inf {xn } (t) = inf Σ(t − s) + inf {xn (s)} n

n

0≤s≤t

=

inf inf {Σ(t − s) + xn (s)}

0≤s≤t n

= inf inf {Σ(t − s) + xn (s)} n 0≤s≤t

= inf {CΣ (xn )(t)} . n

To show that PL is upper semi-continuous, we proceed in two steps. Let x = inf n {xn }. We first note that PL inf {xn } = PL (x ) ≤ inf {PL (xn )} n

n

≤ xn for any n and is a wide-sense increasing function. We next show that the converse because inequality also holds. We first assume that there is some m such that xm = x , namely that the infimum is actually a minimum. Then inf {PL (xn )} ≤ PL (xm ) = PL (x ) . x

PL

n

We next suppose that there is no integer n such that xn = x . Then for any ε > 0, there is an integer m such that 0 < xm − x < ε. Therefore inf {PL (xn )} ≤ PL (xm ) ≤ PL (x + ε) . n

Since the above inequality is true for any ε > 0, and since P L is a right-continuous function, it implies that inf {PL (xn )} ≤ PL (x ) = PL inf {xn } . n

n

This concludes the proof. On the other hand, Dσ is not upper semi-continuous, because its application to an inf would involve the three operations sup, inf and +, which do not commute, as we have seen at the end of the previous chapter. It is easy to show that if Π1 and Π2 are upper semi-continuous, so are Π1 ∧ Π2 and Π1 ◦ Π2 . The dual definition of upper semi-continuity is that of lower semi-continuity, which is defined as follows. D EFINITION 4.1.9 (L OWER SEMI - CONTINUOUS OPERATOR ). Operator Π is lower semi-continuous if for any (finite or infinite) set of functions or sequences {xn }, xn ∈ G J ,

(4.2) Π sup{xn } = sup {Π(xn )} . n

n

It is easy to check that Dσ is lower semi-continuous, unlike other operators, except ST which is also lower semi-continuous.

4.1.5

I SOTONE O PERATORS

D EFINITION 4.1.10 (I SOTONE OPERATOR ). Operator Π is isotone if x1 ≤ x2 always implies Π(x1 ) ≤ Π(x2 ). All upper semi-continuous operators are isotone. Indeed, if x1 ≤ x2 , then x1 ∧ x2 = x1 and since Π is upper semi-continuous, Π(x1 ) = Π(x1 ∧ x2 ) = Π(x1 ) ∧ Π(x2 ) ≤ Π(x2 ). Likewise, all lower semi-continuous operators are isotone. Indeed, if x1 ≤ x2 , then x1 ∨ x2 = x2 and since Π is lower semi-continuous, Π(x1 ) ≤ Π(x1 ) ∨ Π(x2 ) = Π(x1 ∨ x2 ) = Π(x2 ).


136

4.1.6

L INEAR O PERATORS

In classical system theory on (R, +, ×), a system Π is linear if its output to a linear combination of inputs is the linear combination of the outputs to each particular input. In other words, Π is linear if for any (finite or infinite) set of inputs {xi }, and for any constant k ∈ R, xi = Π(xi ) Π i

i

and for any input x and any constant k ∈ R, Π (k · x) = k · Π(x). The extension to min-plus system theory is straightforward. The first property being replaced by that of upper semi-continuity, a min-plus linear operator is thus defined as an upper semi-continuous operator that has the following property (“multiplication” by a constant): D EFINITION 4.1.11 (M IN - PLUS LINEAR OPERATOR ). Operator Π is min-plus linear if it is upper semicontinuous and if for any x ∈ G J and for any k ≥ 0, Π (x + k) = Π (x) + k.

(4.3)

One can easily check that Cσ , CΣ , hσ and ST are min-plus linear, unlike Dσ and PL . Dσ is not linear because it is not upper semi-continuous, and PL is not linear because it fails to verify (4.3). In classical linear theory, a linear system is represented by its impulse response h(t, s), which is defined as the output of the system when the input is the Dirac function. The output of such a system can be expressed as ∞ h(t, s)x(s)ds Π(x)(t) = −∞

Its straightforward extension in Min-plus system theory is provided by the following theorem [28]. To prove this theorem in the vector case, we need first to extend the burst delay function introduced in Definition 3.1.2, to allow negative values of the delay, namely, the value T in 0 if t ≤ T δT (t) = ∞ if t > T, is now taking values in R. We also introduce the following matrix DT ∈ G J × G J . D EFINITION 4.1.12 (S HIFT MATRIX ). The shift matrix is defined by ⎡ ∞ ∞ ··· ∞ δT (t) ⎢ ∞ δ (t) ∞ T ⎢ ⎢ .. . . ∞ δT (t) . . DT (t) = ⎢ ⎢ ∞ ⎢ .. .. . . ⎣ . . . ∞ ∞ ··· ∞ δT (t)

⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦

for some T ∈ R. T HEOREM 4.1.1 (M IN - PLUS IMPULSE RESPONSE ). Π is a min-plus linear operator if and only if there is a unique matrix H ∈ G˜J (called the impulse response), such that for any x ∈ G J and any t ∈ R, Π(x)(t) = inf {H(t, s) + x(s)} . s∈R

(4.4)


137

P ROOF : If (4.4) holds, one immediately sees that Π is upper semi-continuous and verifies (4.3), and therefore is min-plus linear. Π maps G J to G J because H ∈ G˜J . Suppose next that Π is min-plus linear, and let us prove that there is a unique matrix H(t, s) ∈ G˜J such that (4.4) holds. Let us first note that Ds (t) + x(s) = x(s) for any s ≥ t. Since x ∈ G J , we have inf {Ds (t) + x(s)} = inf {x(s)} = x(t).

s≥t

s≥t

On the other hand, all entries of Ds (t) are infinite for s < t. We have therefore that inf {Ds (t) + x(s)} = ∞

s t to hσ is

H(t, s) =

and to ST is

σ(t) − σ(s) 0

if s ≤ t , if s > t

H(t, s) = DT (t − s).

In fact the introduction of the shift matrix allows us to write the shift operator as a min-plus convolution: ST = CDT if T ≥ 0. Let us now compute the impulse response of the compostion of two min-plus linear operators. T HEOREM 4.1.2 (C OMPOSITION OF MIN - PLUS LINEAR OPERATORS ). Let LH and LH be two min-plus linear operators. Then their composition LH ◦ LH is also min-plus linear, and its impulse repsonse denoted by H ◦ H is given by (H ◦ H )(t, s) = inf H(t, u) + H (u, s) . u∈R

P ROOF :

The composition LH ◦ LH applied to some x ∈ G J is

LH (LH (x))(t) = inf H(t, u) + inf H (u, s) + x(s) u s = inf inf H(t, u) + H (u, s) + x(s) u s

= inf inf H(t, s) + H (u, s) + x(s) . s

u


139

We can therefore write LH ◦ LH = LH◦H . Likewise, one easily shows that LH ∧ LH = LH∧H . Finally, let us mention the dual definition of a max-plus linear operator. D EFINITION 4.1.13 (M AX - PLUS LINEAR OPERATOR ). Operator Π is max-plus linear if it is lower semicontinuous and if for any x ∈ G J and for any k ≥ 0, Π (x + k) = Π (x) + k.

(4.7)

Max-plus linear operators can also be represented by their impulse response. T HEOREM 4.1.3 (M AX - PLUS IMPULSE RESPONSE ). Π is a max-plus linear operator if and only if there is a unique matrix H ∈ G˜J (called the impulse response), such that for any x ∈ G J and any t ∈ R, Π(x)(t) = sup {H(t, s) + x(s)} .

(4.8)

s∈R

One can easily check that Dσ and ST are max-plus linear, unlike CΣ , hσ and PL . For example, Dσ (x)(t) can be written as Dσ (x)(t) = sup{x(t + u) − σ(u)} = sup{x(s) − σ(s − t)} = sup{x(s) − σ(s − t)} u≥0

s≥t

s∈R

which has the form (4.8) if H(t, s) = −σ(s − t). Likewise, ST (x)(t) can be written as ST (x) (t) = x(t − T ) = sup{x(s) − D−T (s − t)} s∈R

which has the form (4.8) if H(t, s) = −D−T (s − t).

4.1.7

C AUSAL O PERATORS

A system is causal if its output at time t only depends on its input before time t. D EFINITION 4.1.14 (C AUSAL OPERATOR ). Operator Π is causal if for any t, x1 (s) = x2 (s) for all s ≤ t always implies Π(x1 )(t) = Π(x2 )(t). T HEOREM 4.1.4 (M IN - PLUS CAUSAL LINEAR OPERATOR ). A min-plus linear system with impulse response H is causal if H(t, s) = H(t, t) for s > t. P ROOF :

If H(t, s) = 0 for s > t and if x1 (s) = x2 (s) for all s ≤ t then since x1 , x2 ∈ G J , LH (x1 )(t) =

inf {H(t, s) + x1 (s)}

s∈R

= inf {H(t, s) + x1 (s)} ∧ inf {H(t, s) + x1 (s)} s≤t

s>t

= inf {H(t, s) + x1 (s)} ∧ inf {H(t, t) + x1 (s)} s≤t

= inf {H(t, s) + x1 (s)} s≤t

s>t


140

= inf {H(t, s) + x2 (s)} s≤t

= inf {H(t, s) + x2 (s)} ∧ inf {H(t, t) + x2 (s)} s>t

s≤t

= inf {H(t, s) + x2 (s)} ∧ inf {H(t, s) + x2 (s)} s>t

s≤t

=

inf {H(t, s) + x2 (s)} = LH (x2 )(t).

s∈R

Cσ , CΣ , hσ and PL are causal. ST is causal if and only if T ≥ 0. Dσ is not causal. Indeed if x1 (s) = x2 (s) for all s ≤ t, but that x1 (s) = x2 (s) for all s > t, then Dσ (x1 )(t) = sup {x1 (t + u) − σ(u)} u≥0

= sup {x2 (t + u) − σ(u)} u≥0

= Dσ (x1 )(t)

4.1.8

S HIFT-I NVARIANT O PERATORS

A system is shift-invariant, or time-invariant, if a shift of the input of T time units yields a shift of the output of T time units too. D EFINITION 4.1.15 (S HIFT- INVARIANT OPERATOR ). Operator Π is shift-invariant if it commutes with all shift operators, i.e. if for any x ∈ G and for any T ∈ R Π(ST (x)) = ST (Π(x)). T HEOREM 4.1.5 (S HIFT- INVARIANT MIN - PLUS LINEAR OPERATOR ). Let LH and LH be two min-plus linear, shift-invariant operators. (i) A min-plus linear operator LH is shift-invariant if and only if its impulse response H(t, s) depends only on the difference (t − s). (ii) Two min-plus linear, shift-invariant operators LH and LH commute. If they are also causal, the impulse response of their composition is (H ◦ H )(t, s) = inf H(t − s − u) + H (u) = (H ⊗ H )(t − s). 0≤u≤t−s

P ROOF : (i) Let hj (t, s) and ds,j (t) denote (respectively) the jth column of H(t, s) and of Ds (t). Note that ds,j (t) = Ss (d0,j )(t). Then (4.6) yields that hj (t, s) = Π ds,j (t) = Π Ss (d0,j ) (t) = Ss Π(d0,j ) (t) = Π(d0,j ) (t − s) = hj (t − s, 0) Therefore H(t, s) can be written as a function of a single variable H(t − s). (ii) Because of Theorem 4.1.2, the impulse response of LH ◦ LH is (H ◦ H )(t, s) = inf H(t, u) + H (u, s) . u

Since H(t, u) = H(t − u) and H (u, s) = H (u − s), and setting v = u − s, the latter can be written as (H ◦ H )(t, s) = inf H(t − u) + H (u − s) = inf H(t − s − v) + H (v) . u

v

4.2. CLOSURE OF AN OPERATOR

141

Similarly, the impulse response of LH ◦ LH can be written as (H ◦ H)(t, s) = inf H (t − u) + H(u − s) = inf H(v) + H (t − s − v) u

v

where this time we have set v = t − u. Both impulse responses are identical, which shows that the two operators commute. If they are causal, then their impulse response is infinite for t > s and the two previous relations become (H ◦ H )(t, s) = (H ◦ H)(t, s) = inf H(t − s − v) + H (v) = (H ⊗ H )(t − s). 0≤v≤t

Min-plus convolution CΣ (including of course Cσ and ST ) is therefore shift-invariant. In fact, it follows from this theorem that the only min-plus linear, causal and shift-invariant operator is min-plus convolution. Therefore hσ is not shift-invariant. Min-plus deconvolution is shift-invariant, as Dσ (ST (x))(t) = sup{ST (x)(t + u) − σ(u)} = sup{x(t + u − T ) − σ(u)} u≥0

u≥0

= (x σ)(t − T ) = Dσ (x)(t − T ) = ST (Dσ ) (x)(t). Finally let us mention that PL is not shift-invariant.

4.1.9

I DEMPOTENT O PERATORS

An idempotent operator is an operator whose composition with itself produces the same operator. D EFINITION 4.1.16 (I DEMPOTENT OPERATOR ). Operator Π is idempotent if its self-composition is Π, i.e. if Π ◦ Π = Π. We can easily check that hσ and PL are idempotent. If σ is sub-additive, with σ(0) = 0, then Cσ ◦ Cσ = Cσ , which shows that in this case, Cσ is idempotent too. The same applies to Dσ .

4.2

C LOSURE OF AN O PERATOR

By repeatedly composing a min-plus operator with itself, we obtain the closure of this operator. The formal definition is as follows. D EFINITION 4.2.1 (S UB - ADDITIVE CLOSURE OF AN OPERATOR ). Let Π be a min-plus operator taking G J → G J . Denote Π(n) the operator obtained by composing Π (n − 1) times with itself. By convention, Π(0) = S0 = CD0 , so Π(1) = Π, Π(2) = Π ◦ Π, etc. Then the sub-additive closure of Π, denoted by Π, is defined by

Π = S0 ∧ Π ∧ (Π ◦ Π) ∧ (Π ◦ Π ◦ Π) ∧ . . . = inf Π(n) . (4.9) n≥0

In other words, Π(x) = x ∧ Π(x) ∧ Π(Π(x)) ∧ . . . It is immediate to check that Π does map functions in G J to functions in G J . The next theorem provides the impulse response of the sub-additive closure of a min-plus linear operator. It follows immediately from applying recursively Theorem 4.1.2.


142

T HEOREM 4.2.1 (S UB - ADDITIVE CLOSURE OF A LINEAR OPERATOR ). The impulse response of LH is H(t, s) = inf

inf

n∈N un ,...,u2 ,u1

{H(t, u1 ) + H(u1 , u2 ) + . . . + H(un , s)} .

(4.10)

and LH = LH . For a min-plus linear, shift-invariant and causal operator, (4.10) becomes H(t − s) = inf

inf

n∈N s≤un ≤...≤u2 ≤u1 ≤t

= inf

inf

{H(t − u1 ) + H(u1 − u2 ) + . . . + H(un − s)}

n∈N 0≤vn ≤...≤v2 ≤v1 ≤t−s

{H(t − s − v1 ) + H(v1 − v2 ) + . . . + H(vn )}

= inf {H (n) }(t − s)

(4.11)

n∈N

where H (n) = H ⊗ H ⊗ . . . ⊗ H (n times, n ≥ 1) and H (0) = S0 . In particular, if all entries σij (t) of Σ(t) are sub-additive functions, we find that C Σ = CΣ . In the scalar case, the closure of the min-plus convolution operator Cσ reduces to the min-plus convolution of the sub-additive closure of σ: C σ = Cσ . If σ is a “good” function (i.e., a sub-additive function with σ(0) = 0), then C σ = Cσ . The sub-additive closure of the idempotent operators hσ and PL are easy to compute too. Indeed, since hσ (x) ≤ x and PL (x) ≤ x, hσ = hσ and P L = PL . The following result is easy to prove. We write Π ≤ Π to express that Π(x) ≤ Π (x) for all x ∈ G J . T HEOREM 4.2.2 (S UB - ADDITIVE CLOSURE OF AN ISOTONE OPERATOR ). If Π and Π are two isotone operators, and Π ≤ Π , then Π ≤ Π . Finally, let us conclude this section by computing the closure of the minimum between two operators. T HEOREM 4.2.3 (S UB - ADDITIVE CLOSURE OF Π1 ∧ Π2 ). Let Π1 , Π2 be two isotone operators taking G J → G J . Then Π1 ∧ Π2 = (Π1 ∧ S0 ) ◦ (Π2 ∧ S0 ). (4.12) P ROOF :

(i) Since S0 is the identity operator, Π1 ∧ Π2 = (Π1 ◦ S0 ) ∧ (S0 ◦ Π2 ) ≥ ((Π1 ∧ S0 ) ◦ S0 ) ∧ (S0 ◦ (Π2 ∧ S0 )) ≥ ((Π1 ∧ S0 ) ◦ (Π2 ∧ S0 )) ∧ ((Π1 ∧ S0 ) ◦ (Π2 ∧ S0 )) = (Π1 ∧ S0 ) ◦ (Π2 ∧ S0 ).

Since Π1 and Π2 are isotone, so are Π1 ∧ Π2 and (Π1 ∧ S0 ) ◦ (Π2 ∧ S0 ). Consequently, Theorem 4.2.2 yields that Π1 ∧ Π2 ≥ (Π1 ∧ S0 ) ◦ (Π1 ∧ S0 ). (4.13)

4.2. CLOSURE OF AN OPERATOR

143

(ii) Combining the two inequalities Π1 ∧ S0 ≥ Π1 ∧ Π2 ∧ S0 Π2 ∧ S0 ≥ Π1 ∧ Π2 ∧ S0 we get that (Π1 ∧ S0 ) ◦ (Π1 ∧ S0 ) ≥ (Π1 ∧ Π2 ∧ S0 ) ◦ (Π1 ∧ Π2 ∧ S0 ).

(4.14)

Let us show by induction that ((Π1 ∧ Π2 ) ∧ S0 )(n) = min

0≤k≤n

(Π1 ∧ Π2 )(k) .

Clearly, the claim holds for n = 0, 1. Suppose it is true up to some n ∈ N. Then ((Π1 ∧ Π2 ) ∧ S0 )(n+1) = ((Π1 ∧ Π2 ) ∧ S0 ) ◦ ((Π1 ∧ Π2 ) ∧ S0 )(n)

= ((Π1 ∧ Π2 ) ∧ S0 ) ◦ min (Π1 ∧ Π2 )(k) 0≤k≤n

(k) (k) ∧ S0 ◦ min (Π1 ∧ Π2 ) = (Π1 ∧ Π2 ) ◦ min (Π1 ∧ Π2 ) 0≤k≤n 0≤k≤n

(k) (k) = min (Π1 ∧ Π2 ) ∧ min (Π1 ∧ Π2 ) 1≤k≤n+1 0≤k≤n

(Π1 ∧ Π2 )(k) . = min 0≤k≤n+1

Therefore the claim holds for all n ∈ N, and (((Π1 ∧ Π2 ) ∧ S0 ) ◦ ((Π1 ∧ Π2 ) ∧ S0 ))(n) = ((Π1 ∧ Π2 ) ∧ S0 )(2n)

= min (Π1 ∧ Π2 )(k) . 0≤k≤2n

Consequently,

inf min (Π1 ∧ Π2 )(k) n∈N 0≤k≤2n

= inf (Π1 ∧ Π2 )(k)

(Π1 ∧ Π2 ∧ S0 ) ◦ (Π1 ∧ Π2 ∧ S0 ) =

k∈N

= Π1 ∧ Π2 and combining this result with (4.13) and (4.14), we get (4.12). If one of the two operators is an idempotent operator, we can simplify the previous result a bit more. We will use the following corollary in Chapter 9. C OROLLARY 4.2.1 (S UB - ADDITIVE CLOSURE OF Π1 ∧hM ). Let Π1 be an isotone operator taking F → F, and let M ∈ F. Then Π1 ∧ hM = (hM ◦ Π1 ) ◦ hM . (4.15) P ROOF :

Theorem 4.2.3 yields that Π1 ∧ hM = (Π1 ∧ S0 ) ◦ hM

(4.16)


144

because hM ≤ S0 . The right hand side of (4.16) is the inf over all integers n of ({Π1 ∧ S0 } ◦ hM )(n) which we can expand as {Π1 ∧ S0 } ◦ hM ◦ {Π1 ∧ S0 } ◦ hM ◦ . . . ◦ {Π1 ∧ S0 } ◦ hM . Since hM ◦ {Π1 ∧ S0 } ◦ hM

= {hM ◦ Π1 ◦ hM } ∧ hM = ({hM ◦ Π1 } ∧ S0 ) ◦ hM

= min (hM ◦ Π1 )(q) ◦ hM , 0≤q≤1

the previous expression is equal to

min

0≤q≤n

(hM ◦ Π1 )(q) ◦ hM .

Therefore we can rewrite the right hand side of (4.16) as

(q) ◦ hM min (hM ◦ Π1 ) (Π1 ∧ S0 ) ◦ hM = inf n∈N 0≤q≤n

= inf (hM ◦ Π1 )(q) ◦ hM = (hM ◦ Π1 ) ◦ hM , q∈N

which establishes (4.15). Therefore we can rewrite the right hand side of (4.16) as

(q) ◦ hM min (hM ◦ Π1 ) (Π1 ∧ S0 ) ◦ hM = inf n∈N 0≤q≤n

= hM ◦ inf (hM ◦ Π1 )(q) ◦ hM = hM ◦ (hM ◦ Π1 ) ◦ hM , q∈N

which establishes (4.15). The dual of super-additive closure is that of super-additive closure, defined as follows. D EFINITION 4.2.2 (S UPER - ADDITIVE CLOSURE OF AN OPERATOR ). Let Π be an operator taking G J → G J . The super-additive closure of Π, denoted by Π, is defined by

(4.17) Π = S0 ∨ Π ∨ (Π ◦ Π) ∨ (Π ◦ Π ◦ Π) ∨ . . . = sup Π(n) . n≥0

4.3 4.3.1

F IXED P OINT E QUATION (S PACE M ETHOD ) M AIN T HEOREM

We now have the tools to solve an important problem of network calculus, which has some analogy with ordinary differential equations in conventional system theory. The latter problem reads as follows: let Π be an operator from RJ to RJ , and let a ∈ RJ . What is then the solution x(t) to the differential equation dx (t) = Π(x)(t) dt

(4.18)

4.3. FIXED POINT EQUATION (SPACE METHOD)

145

with the inital condition x(0) = a.

(4.19)

Here Π is an operator taking G J → G J , and a ∈ G J . The problem is now to find the largest function x(t) ∈ G J , which verifies the recursive inequality x(t) ≤ Π(x)(t)

(4.20)

x(t) ≤ a(t).

(4.21)

and the initial condition The differences are however important: first we have inequalities instead of equalities, and second, contrary to (4.18), (4.20) does not describe the evolution of the trajectory x(t) with time t, starting from a fixed point a, but the successive iteration of Π on the whole trajectory x(t), starting from a fixed, given function a(t) ∈ G J . The following theorem provides the solution this problem, under weak, technical assumptions that are almost always met. T HEOREM 4.3.1 (S PACE METHOD ). Let Π be an upper semi-continuous operator taking G J → G J . For any fixed function a ∈ G J , the problem x ≤ a ∧ Π(x) (4.22) has one maximum solution in G J , given by x = Π(a). The theorem is proven in [28]. We give here a direct proof that does not have the pre-requisites in [28]. It is based on a fixed point argument. We call the application of this theorem “Space method”, because the iterated variable is not time t (as in the “Time method” described shortly later) but the full sequence x itself. The theorem applies therefore indifferently whether t ∈ Z or t ∈ R. P ROOF : (i) Let us first show that Π(a) is a solution of (4.22). Consider the sequence {xn } of decreasing sequences defined by x0 = a xn+1 = xn ∧ Π(xn ),

n ≥ 0.

Then one checks that x = inf {xn } n≥0

is a solution to (4.22) because

x

≤ x0 = a and because Π is upper-semi-continuous so that

Π(x ) = Π( inf {xn }) = inf {Π(xn )} ≥ inf {xn+1 } ≥ inf {xn } = x . n≥0

n≥0

n≥0

n≥0

Now, one easily checks that xn = inf 0≤m≤n {Π(m) (a)}, so x = inf {xn } = inf n≥0

inf {Π(m) (a)} = inf {Π(n) (a)} = Π(a).

n≥0 0≤m≤n

n≥0

This also shows that x ∈ G J . (ii) Let x be a solution of (4.22). Then x ≤ a and since Π is isotone, Π(x) ≤ Π(a). From (4.22), x ≤ Π(x), so that x ≤ Π(a). Suppose that for some n ≥ 1, we have shown that x ≤ Π(n−1) (a). Then as x ≤ Π(x) and as Π is isotone, it yields that x ≤ Π(n) (a). Therefore x ≤ inf n≥0 {Π(n) (a)} = Π(a), which shows that x = Π(a) is the maximal solution. Similarly, we have the following result in Max-plus algebra.

146


T HEOREM 4.3.2 (D UAL SPACE METHOD ). Let Π be a lower semi-continuous operator taking G J → G J . For any fixed function a ∈ G J , the problem x ≥ a ∨ Π(x)

(4.23)

has one minimum solution, given by x = Π(a).

4.3.2

E XAMPLES OF A PPLICATION

Let us now apply this theorem to five particular examples. We will first revisit the input-output characterization of the greedy shaper of Section 1.5.2, and of the variable capacity node described at the end of Section 1.3.2. Next we will apply it to two window flow control problems (with a fixed length window). Finally, we will revisit the variable length packet greedy shaper of Section 1.7.4. I NPUT-O UTPUT C HARACTERIZATION OF G REEDY S HAPERS Remember that a greedy shaper is a system that delays input bits in a buffer, whenever sending a bit would violate the constraint σ, but outputs them as soon as possible otherwise. If R is the input flow, the output is thus the maximal function x ∈ F satisfying the set of inequalities (1.13), which we can recast as x ≤ R ∧ Cσ (x). It is thus given by R∗ = Cσ = Cσ (x) = σ ⊗ x. If σ is a “good” function, one therefore retrieves the main result of Theorem 1.5.1. I NPUT-O UTPUT C HARACTERIZATION OF VARIABLE C APACITY N ODES The variable capacity node was introduced at the end of Section 1.3.2, where the variable capacity is modeled by a cumulative function M (t), where M (t) is the total capacity available to the flow between times 0 and t. If m(t) is the instantaneous capacity available to the flow at time t, then M (t) is the primitive of this function. In other words, if t ∈ R, t

m(s)ds

M (t) =

(4.24)

0

and if t ∈ Z the integral is replaced by a sum on s. If R is the input flow and x is the output flow of the variable capacity node, then the variable capacity constraint imposes that for all 0 ≤ s ≤ t x(t) − x(s) ≤ M (t) − M (s), which we can recast using the idempotent operator hM as x ≤ hM (x).

(4.25)

x ≤ R.

(4.26)

On the other hand, the system is causal, so that

The output of the variable capacity node is therefore the maximal solution of system (4.25) and (4.26). It is thus given by R∗ (t) = hM (R)(t) = hM (R)(t) = inf {M (t) − M (s) + R(s)} 0≤s≤t

because the sub-additive closure of an idempotent operator is the operator itself, as we have seen in the previous section.

4.3. FIXED POINT EQUATION (SPACE METHOD)

147

S TATIC WINDOW FLOW CONTROL – EXAMPLE 1 Let us now consider an example of a feedback system. This example is found independently in [10] and [68, 2]. A data flow a(t) is fed via a window flow controller to a network offering a service curve β. The window flow controller limits the amount of data admitted into the network in such a way that the total backlog is less than or equal to W , where W > 0 (the window size) is a fixed number (Figure 4.1).

controller network a(t)

x(t) y(t)

Figure 4.1: Static window flow control, from [10] or [68] Call x(t) the flow admitted to the network, and y(t) the output. The definition of the controller means that x(t) is the maximum solution to x(t) ≤ a(t) (4.27) x(t) ≤ y(t) + W We do not know the mapping Π : x → y = Π(x), but we assume that Π is isotone, and we assume that y(t) ≥ (β ⊗ x)(t), which can be recast as Π(x) ≥ Cβ (x).

(4.28)

x ≤ a ∧ {Π(x) + W } ,

(4.29)

We also recast System (4.27) as and direclty apply Theorem 4.3.1 to derive that the maximum solution is x = (Π + W )(a). Since Π is isotone, so is Π + W . Therefore, because of (4.28) and applying Theorem 4.2.2, we get that x = (Π + W )(a) ≥ (Cβ + W )(a).

(4.30)

Because of Theorem 4.2.1, (Cβ + W )(a) = C β+W (a) = Cβ+W (a) = (β + W ) ⊗ a. Combining this relationship with (4.30) we have that y ≥ β ⊗ x ≥ β ⊗ (β + W ) ⊗ a = β ⊗ (β + W ) (a), which shows that the complete, closed-loop system of Figure 4.1 offers to flow a a service curve [10] βwfc1 = β ⊗ (β + W ).

(4.31)

For example, if β = βR,T then the service curve of the closed-loop system is the function represented on Figure 4.2. When RT ≤ W , the window does not add any restriction on the service guarantee offered by the open-loop system, as in this case βwfc1 = β. If RT > W on the other hand, the service curve is smaller than the open-loop service curve.


148

βwfc1(t)

βwfc1(t) = β(t) = R[t-T]+

R W

R t

T

W

t T 2T 3T 4T

Case 1: RT ≤ W

Case 2: RT > W

Figure 4.2: The service curve βwfc1 of the closed-loop system with static window flow control, when the service curve of the open loop system is βR,T with RT ≤ W (left) and RT > W (right). S TATIC WINDOW FLOW CONTROL – EXAMPLE 2 Let us extend the window flow control model to account for the existence of background traffic, which constraints the input traffic rate at time t, dx/dt(t) (if t ∈ R) or x(t) − x(t − 1) (if t ∈ Z), to be less that some given rate m(t). Let M (t) denote the primitive of this prescribed rate function. Then the rate constraint on x becomes (4.25). Function M (t) is not known, but we assume that there is some function γ ∈ F such that M (t) − M (s) ≥ γ(t − s) for any 0 ≤ s ≤ t, which we can recast as

hM ≥ Cγ .

(4.32)

This is used in [47] to derive a service curve offered by the complete system to the incoming flow x, which we shall also compute now by applying Theorem 4.3.1. With the additional constraint (4.25), one has to compute the maximal solution of x ≤ a ∧ {Π(x) + W } ∧ hM (x),

(4.33)

x = ({Π + W } ∧ hM )(a).

(4.34)

which is As in the previous subsection, we do not know Π but we assume that it is isotone and that Π ≥ Cβ . We also know that hM ≥ Cγ . A first approach to get a service curve for y, is to compute a lower bound of the right hand side of (4.34) by time-invariant linear operators, which commute as we have seen earlier in this chapter. We get {Π + W } ∧ hM ≥ {Cβ + W } ∧ Cγ = C{β+W }∧γ , and therefore (4.34) becomes x ≥ C {β+W }∧γ (a) = C{β+W }∧γ (a) = ({β + W } ∧ γ) ⊗ a. Because of Theorem 3.1.11, {β + W } ∧ γ = (β + W ) ⊗ γ so that

y ≥ β ⊗ x ≥ β ⊗ (β + W ) ⊗ γ ⊗ a

4.4. FIXED POINT EQUATION (TIME METHOD)

149

and thus a service curve for flow a is β ⊗ (β + W ) ⊗ γ.

(4.35)

Unfortunately, this service curve can be quite useless. For example, if for some T > 0, γ(t) = 0 for 0 ≤ t ≤ T , then γ(t) = 0 for all t ≥ 0, and so the service curve is zero. A better bound is obtained by differing the lower bounding of hM by the time-invariant operator Cγ after having used the idempotency property in the computation of the sub-additive closure of the right hand side of (4.34), via Corollary 4.2.1. Indeed, this corollary allows us to replace (4.34) by x = (hM ◦ (Π + W )) ◦ hM (a). Now we can bound hM below by Cγ to obtain (hM ◦ (Π + W )) ◦ hM

≥ (Cγ ◦ Cβ+W ) ◦ Cγ = C γ⊗(β+W ) ◦ Cγ = Cβ⊗γ+W ◦ Cγ = Cγ⊗(β⊗γ+W ) .

We obtain a better service curve than by our initial approach, where we had directly replaced hM by Cgamma : βwfc2 = β ⊗ γ ⊗ (β ⊗ γ + W ).

(4.36)

is a better service curve than (4.35). For example, if β = βR,T and γ = βR ,T , with R > R and W < R (T + T ), then the service curve of the closed-loop system is the function represented on Figure 4.3. PACKETIZED GREEDY SHAPER Our last example in this chapter is the packetized greedy shaper introduced in Section 1.7.4. It amounts to computing the maximum solution to the problem x ≤ R ∧ PL (x) ∧ Cσ (x) where R is the input flow, σ is a “good” function and L is a given sequence of cumulative packet lengths. We can apply Theorem 4.3.1 and next Theorem 4.2.2 to obtain x = PL ∧ Cσ (R) = PL ◦ Cσ (R) which is precisely the result of Theorem 1.7.4.

4.4

F IXED P OINT E QUATION (T IME M ETHOD )

We conclude this chapter by another version of Theorem 4.3.1 that applies only to the disrete-time setting. It amounts to compute the maximum solution x = Π(a) of (4.22) by iterating on time t instead of interatively applying operator Π to the full trajectory a(t). We call this method the “time method” (see also [11]). It is valid under stronger assumptions than the space method, as we require here that operator Π be min-plus linear.


150

γ(t) = R’[t-T’]+

β(t) = R[t-T]+

R

R’ T’

t

T

t

βwfc2(t)

(β ⊗ γ)(t)

R’ R’ t

W

t

T+T’

T+T’

Figure 4.3: The service curve βwfc2 of the closed-loop system with window flow control (bottom right), when the service curve of the open loop system is β = βR,T (top left) and when γ = βR ,T (top right), with R > R and W < R (T + T ). T HEOREM 4.4.1. Let Π = LH be a min-plus linear operator taking F J → F J , with impulse response H ∈ F˜ J . For any fixed function a ∈ F J , the problem x ≤ a ∧ LH (x)

(4.37)

has one maximum solution, given by x (0) = a(0) x (t) = a(0) ∧

inf

{H(t, u) + x (u)}.

0≤u≤t−1

P ROOF : Note that the existence of a maximum solution is given by Theorem 4.3.1. Define x by the recursion in the Theorem. As H ∈ F˜ J it follows easily by induction that x is a solution to problem (4.37). Conversely, for any solution x, x(0) ≤ a(0) = x (0) and if x(u) ≤ x (u) for all 0 ≤ u ≤ t − 1, it follows that x(t) ≤ x (t) which shows that x is the maximal solution.

4.5

C ONCLUSION

This chapter has introduced min-plus and max-plus operators, and discussed their properties, which are summarized in Table 4.5. The central result of this chapter, which will be applied in the next chapters, is Theorem 4.3.1, which enables us to compute the maximal solution of a set of inqualities involving the iterative application of an upper semi-continuous operator.

4.5. CONCLUSION

151

Operator Upper semi-continuous Lower semi-continuous Isotone Min-plus linear Max-plus linear Causal Shift-invariant Idempotent

Cσ yes no yes yes no yes yes no (2)

Dσ no yes yes no yes no yes no (2)

ST yes yes yes yes yes yes (1) yes no (3)

hσ yes no yes yes no yes no yes

PL yes no yes no no yes no yes

(1) (if T ≥ 0) (2) (unless σ is a ‘good’ function) (3) (unless T = 0) Table 4.1: A summary of properties of some common operators

152


PART III

A S ECOND C OURSE IN N ETWORK C ALCULUS

153

C HAPTER 5

O PTIMAL M ULTIMEDIA S MOOTHING In this chapter we apply network calculus to smooth multimedia data over a network offering reservation based services, such as ATM or RSVP/IP, for which we know one minimal service curve. One approach to stream video is to act on the quantization levels at the encoder output: this is called rate control, see e.g. [26]. Another approach is to smooth the video stream, using a smoother fed by the encoder, see e.g. [69, 72, 59]. In this chapter, we deal with this second approach. A number of smoothing algorithms have been proposed to optimize various performance metrics, such as peak bandwidth requirements, variability of transmission rates, number of rate changes, client buffer size [29]. With network calculus, we are able to compute the minimal client buffer size required given a maximal peak rate, or even a more complex (VBR) smoothing curve. We can also compute the minimal peak rate required given a given client buffer size. We will see that the scheduling algorithm that must be implemented to reach these bounds is not unique, and we will determine the full set of video transmission schedules that minimize these resources and achieve these optimal bounds.

5.1

P ROBLEM S ETTING

A video stream stored on the server disk is directly delivered to the client, through the network, as shown on Figure 5.1. At the sender side, a smoothing device reads the encoded video stream R(t) and sends a stream x(t) that must conform to an arrival curve σ, which we assume to be a ‘good’ function, i.e. is sub-additive and such that σ(0) = 0. The simplest and most popular smoothing curve in practice is a constant rate curve (or equivalently, a peak rate constraint) σ = λr for some r > 0. We take the transmission start as origin of time: this implies that x(t) = 0 for t ≤ 0. At the receiver side, the video stream R will be played back after D units of times, the playback delay: the output of the decoding buffer B must therefore be R(t − D). The network offers a guaranteed service to the flow x. If y denotes the output flow, it is not possible, in general, to express y as a function of x. However we assume that the service guarantee can be expressed by a service curve β. For example, as we have seen in Chapter 1, the IETF assumes that RSVP routers offer a rate-latency service curve β of the form βL,C (t) = C[t − L]+ = max{0, C(t − L)}. Another example is a network which is completely transparent to the flow (i.e. which does not incur any jitter to the flow nor rate limitation, even if it can introduce a fixed delay, which we ignore in this chapter as we can always take it into account separately). We speak of a null network. It offers a service curve β(t) = δ0 (t). 155

CHAPTER 5. OPTIMAL MULTIMEDIA SMOOTHING

156

Client video display

Video Server Network σ R(t+d)

x(t) Smoother

β

B y(t)

R(t-D) Client playback buffer

Figure 5.1: Video smoothing over a single network. To keep mathematical manipulations simple, we assume that the encoding buffer size is large enough to contain the full data stream. On the other hand, the receiver (decoding) buffer is a much more scarce resource. Its finite size is denoted by B. As the stream is pre-recorded and stored in the video server, it allows the smoother to prefetch and send some of the data before schedule. We suppose that the smoother is able to look ahead data for up to d time units ahead. This look-ahead delay can take values ranging from zero (in the most restrictive case where no prefetching is possible) up to the length of the full stream. The sum of the look-ahead delay and playback delay is called the total delay, and is denoted by T : T = D + d. These constraints are described more mathematically in Section 5.2. We will then apply Theorem 4.3.1 to solve the following problems: (i) we first compute, in Section 5.3, the minimal requirements on the playback delay D, on the look-ahead delay d, and on the client buffer size B guaranteeing a lossless transmission for given smoothing and service curves σ and β. (ii) we then compute, in Section 5.4, all scheduling strategies at the smoother that will achieve transmission in the parameter setting computed in Section 5.3. We call the resulting scheduling “optimal smoothing”. (iii) in the CBR case (σ = λr ), for a given rate r and for a rate-latency service curve (β = βL,C ), we will obtain, in Section 5.5, closed-form expressions of the minimal values of D, T = D + d and B required for lossless smoothing. We will also solve the dual problem of computing the minimal rate r needed to deliver video for a given playback delay D, look-ahead delay d and client buffer size B. We will then compare optimal smoothing with greedy shaping in Section 5.6 and with separate delay equalization in Section 5.7. Finally, we will repeat problems (i) and (iii) when intermediate caching is allowed between a backbone network and an access network.

5.2

C ONSTRAINTS I MPOSED BY L OSSLESS S MOOTHING

We can now formalize the constraints that completely define the smoothing problem illustrated on Figure 5.1). • Flow x ∈ F: As mentioned above, the chosen origin of time is such that x(t) = 0 for t ≤ 0, or equivalently x(t) ≤ δ0 (t). (5.1) • Smoothness constraint: Flow x is constrained by an arrival curve σ(·). This means that for all t ≥ 0 x(t) ≤ (x ⊗ σ)(t) = Cσ (x)(t).

(5.2)

5.3. MINIMAL REQUIREMENTS ON DELAYS AND PLAYBACK BUFFER

157

• Playback delay constraint (no playback buffer underflow): The data is read out from the playback buffer after D unit of times at a rate given by R(t − D). This implies that y(t) ≥ R(t − D). However we do not know the exact expression of y as a function of x. All we know is that the network guarantees a service curve β, namely that y(t) ≥ (x ⊗ β)(t). The output flow may therefore be as low as (x ⊗ β)(t), and hence we can replace y in the previous inequality to obtain (x ⊗ β)(t) ≥ R(t − D). Using Rule 14 in Theorem 3.1.12, we can recast this latter inequality as x(t) ≥ (R β)(t − D) = Dβ (R)(t − D)

(5.3)

for all t ≥ 0. • Playback buffer constraint (no playback buffer overflow): The size of the playback buffer is limited to B, and to prevent any overflow of the buffer, we must impose that y(t) − R(t − D) ≤ B for all t ≥ 0. Again, we do not know the exact value of y, but we know that it can be as high as x, but not higher, because the network is a causal system. Therefore the constraint becomes, for all t ≥ 0, x(t) ≤ R(t − D) + B.

(5.4)

• Look-ahead delay constraint: We suppose that the encoder can prefetch data from the server up to d time units ahead, which translates in the following inequality: x(t) ≤ R(t + d).

5.3

(5.5)

M INIMAL R EQUIREMENTS ON D ELAYS AND P LAYBACK B UFFER

Inequalities (5.1) to (5.5) can be recast as two sets of inequalities as follows: x(t) ≤ δ0 (t) ∧ R(t + d) ∧ {R(t − D) + B} ∧ Cσ (x)(t)

(5.6)

x(t) ≥ (R β)(t − D).

(5.7)

There is a solution x to the smoothing problem if and only if it simultaneously verifies (5.6) and (5.7). This is equivalent to requiring that the maximal solution of (5.6) is larger than the right hand side of (5.7) for all t. Let us first compute the maximal solution of (5.6). Inequality (5.6) has the form x ≤ a ∧ Cσ (x)

(5.8)

a(t) = δ0 (t) ∧ R(t + d) ∧ {R(t − D) + B}.

(5.9)

where

We can thus apply Theorem 4.3.1 to compute the unique maximal solution of (5.8), which is xmax = Cσ (a) = σ ⊗ a because σ is a ‘good’ function. Replacing a by its expression in (5.9), we compute that the maximal solution of (5.6) is xmax (t) = σ(t) ∧ {(σ ⊗ R)(t + d)} ∧ {(σ ⊗ R)(t − D) + B} .

(5.10)

We are now able to compute the smallest values of the playback delay D, of the total delay T and of the playback buffer B ensuring the existence of a solution to the smoothing problem, thanks to following theorem. The requirement on d for reaching the smallest value of D is therefore d = T − D.


158

T HEOREM 5.3.1 (R EQUIREMENTS FOR OPTIMAL SMOOTHING ). The smallest values of D, T and B ensuring a lossless smoothing to a ‘good’ curve σ through a network offering a service curve β are Dmin = h(R, (β ⊗ σ)) = inf {t ≥ 0 : (R (β ⊗ σ))(−t) ≤ 0}

(5.11)

Tmin = h((R R), (β ⊗ σ))

(5.12)

= inf {t ≥ 0 : ((R R) (β ⊗ σ))(−t) ≤ 0} Bmin = v((R R), (β ⊗ σ)) = ((R R) (β ⊗ σ))(0).

(5.13)

where h and v denote respectively the horizontal and vertical distances given by Definition 3.1.15. P ROOF : The set of inequalities (5.6) and (5.7) has a solution if, and only if, the maximal solution of (5.6) is larger or equal to the right hand side of (5.7) at all times. This amounts to impose that for all t ∈ R (R β)(t − D) − σ(t) ≤ 0 (R β)(t − D) − (σ ⊗ R)(t + d) ≤ 0 (R β)(t − D) − (σ ⊗ R)(t − D) ≤ B. Using the deconvolution operator and its properties, the latter three inequalities can be recast as (R (β ⊗ σ))(−D) ≤ 0 ((R R) (β ⊗ σ)) (−T ) ≤ 0 ((R R) (β ⊗ σ)) (0) ≤ B. The minimal values of D, T and B satisfying these three inequalities are given by (5.11), (5.12) and (5.13). These three inequalities are therefore the necessary and sufficient conditions ensuring the existence of a solution to the smoothing problem.

5.4

O PTIMAL S MOOTHING S TRATEGIES

An optimal smoothing strategy is a solution x(t) to the lossless smoothing problem where D, T = D + d and B take their minimal value given by Theorem 5.3.1. The previous section shows that there exists at least one optimal solution, namely (5.10). It is however not the only one, as we will see in this section.

5.4.1

M AXIMAL S OLUTION

The maximal solution (5.10) requires only the evaluation of an infimum at time t over the past values of R and over the future values of R up to time t + dmin , with dmin = Tmin − Dmin . Of course, we need the knowledge of the traffic trace R(t) to dimension Dmin , dmin and Bmin . However, once we have these values, we do not need the full stream for the computation of the smoothed input to the network.

5.4.2

M INIMAL S OLUTION

To compute the minimal solution, we reformulate the lossless smoothing problem slightly differently. Because of Rule 14 of Theorem 3.1.12, an inequality equivalent to (5.2) is x(t) ≥ (x σ)(t) = Dσ (x)(t).

(5.14)

5.5. OPTIMAL CONSTANT RATE SMOOTHING

159

We use this equivalence to replace the set of inequalities (5.6) and (5.7) by the equivalent set x(t) ≤ δ0 (t) ∧ R(t + d) ∧ {R(t − D) + B} (5.15) x(t) ≥ (R β)(t − D) ∨ Dσ (x)(t).

(5.16)

One can then apply Theorem 4.3.2 to compute the minimal solution of (5.16), which is xmin = Dσ (b) = b σ where b(t) = δ0 (t) ∧ R(t + d) ∧ {R(t − D) + B}, because σ is a ‘good’ function. Eliminating b from these expressions, we compute that the minimal solution is xmin (t) = (R (β ⊗ σ))(t − D),

(5.17)

and compute the constraints on d, D and B ensuring that it verifies (5.15): one would get the very same values of Dmin , Tmin and Bmin given by (5.11) (5.12) and (5.13). It does achieve the values of Dmin and Bmin given by (5.11) and (5.13), but requires nevertheless the evaluation, at time t, of a supremum over all values of R up to the end of the trace, contrary to the maximal solution (5.10). Min-plus deconvolution can however be represented in the time inverted domain by a minplus convolution, as we have seen in Section 3.1.10. As the duration of the pre-recorded stream is usually known, the complexity of computing a min-plus deconvolution can thus be reduced to that of computing a convolution.

5.4.3

S ET OF O PTIMAL S OLUTIONS

Any function x ∈ F such that

xmin ≤ x ≤ xmax

and x≤x⊗σ is therefore also a solution to the lossless smoothing problem, for the same minimal values of the playback delay, look-ahead delay and client buffer size. This gives the set of all solutions. A particular solution among these can be selected to further minimize another metric, such as the ones discussed in [29], e.g. number of rate changes or rate variability. The top of Figure 5.2 shows, for a synthetic trace R(t), the maximal solution (5.10) for a CBR smoothing curve σ(t) = λr (t) and a service curve β(t) = δ0 (t), whereas the bottom shows the minimal solution (5.17). Figure 5.3 shows the same solutions on a single plot, for the MPEG trace R(t) of the top of Figure 1.2.4 representing the number of packet arrivals per time slot of 40 ms corresponding to a MPEG-2 encoded video when the packet size is 416 bytes for each packet. An example of VBR smoothing on the same MPEG trace is shown on Figure 5.4, with a smoothing curve derived from the T-SPEC field, which is given by σ = γP,M ∧ γr,b , where M is the maximum packet size (here M = 416 Bytes), P the peak rate, r the sustainable rate and b the burst tolerance. Here we roughly have P = 560 kBytes/sec, r = 330 kBytes/sec and b = 140 kBytes The service curve is a rate-latency curve βL,C with L = 1 second and r = 370 kBytes/sec. The two traces have the same envelope, thus the same minimum buffer requirement (here, 928kBytes). However the second trace has its bursts later, thus, has a smaller minimum playback delay (D2 = 2.05s versus D1 = 2.81s).

5.5

O PTIMAL C ONSTANT R ATE S MOOTHING

Let us compute the above values in the case of a constant rate (CBR) smoothing curve σ(t) = λr (t) = rt (with t ≥ 0) and a rate-latency service curve of the network β(t) = βL,C (t) = C[t − L]+ . We assume that


160

σ (t)=rt R(t)

xmax(t)

Bmin R(t+dmin ) R(t-Dmin )

dmin Dmin

t

R(t)

xmin(t)

Bmin R(t+dmin ) R(t-Dmin )

t

dmin Dmin

Figure 5.2: In bold, the maximal solution (top figure) and minimal solution (bottom figure) to the CBR smoothing problem with a null network. r < C, the less interesting case where r ≥ C being handled similarly. We will often use the decomposition of a rate-latency function as the min-plus convolution of a pure delay function, with a constant rate function: βL,C = δL ⊗ λC . We will also use the following lemma. L EMMA 5.5.1. If f ∈ F, h(f, βL,C ) = L +

1 (f λC )(0). C

(5.18)

5.5. OPTIMAL CONSTANT RATE SMOOTHING

161

d min D min

5000

cumulative flow [Kbytes]

4500

xmax (t)

4000

x min (t)

3500 3000 2500

R(t + dmin ) R(t)

2000 1500

R(t - Dmin )

1000

Bmin

500

t

0 0

100

200

300

400

500

frame number

Figure 5.3: In bold, the maximal and minimal solutions to the CBR smoothing problem of an MPEG trace with a null network. A frame is generated every 40 msec. 7 0 6 0

1 0 0 0 0

5 0

D 1

8 0 0 0

4 0

6 0 0 0

3 0 2 0

4 0 0 0

1 0

2 0 0 0 1 0 0

2 0 0

3 0 0

4 0 0

1 0 0

2 0 0

3 0 0

4 0 0

3 0 0

4 0 0

7 0 6 0

1 0 0 0 0

5 0

D 2

8 0 0 0

4 0

6 0 0 0

3 0 2 0

4 0 0 0

1 0

2 0 0 0 1 0 0

2 0 0

3 0 0

4 0 0

1 0 0

2 0 0

Figure 5.4: Two MPEG traces with the same arrival curve (left). The corresponding playback delays D1 and D2 are the horizontal deviations between the cumulative flows R(t) and function σ ⊗ β (right).


162 P ROOF :

As f (t) = 0 for t ≤ 0 and as βL,C = δL ⊗ λC , we can write for any t ≥ 0 (f βL,C )(−t) = sup{f (u − t) − (δL ⊗ λC )(u)} u≥0

= sup{f (u − t) − λC (u − L)} u≥0

=

sup {f (v) − λC (v + t − L)}

v≥−t

= sup{f (v) − λC (v + t − L)} v≥0

= sup{f (v) − λC (v)} − C(t − L) v≥0

= (f λC )(0) − Ct + CL, from which we deduce the smallest value of t making the left-hand side of this equation non-positive is given by (5.18). In the particular CBR case, the optimal values (5.11), (5.12) and (5.13) become the following ones. T HEOREM 5.5.1 (R EQUIREMENTS FOR CBR OPTIMAL SMOOTHING ). If σ = λr and β = βL,C with r < C, the smallest values of D, of T and of B are 1 Dmin = L + (R λr )(0) r 1 Tmin = L + ((R R) λr )(0) r Bmin = ((R R) λr ))(L) ≤ rTmin . P ROOF :

(5.19) (5.20) (5.21)

To establish (5.19) and (5.20), we note that R and (R R) ∈ F. Since r < C β ⊗ σ = βL,C ⊗ λr = δL ⊗ λC ⊗ λr = δL ⊗ λr = βL,r

so that we can apply Lemma 5.5.1 with f = R and f = (R R), respectively. To establish (5.21), we develop (5.13) as follows ((R R) (β ⊗ σ))(0) = ((R R) (δL ⊗ λr ))(0) = sup{(R R)(u) − λr (u − L)} u≥0

= ((R R) λr )(L) = sup {(R R)(u) − λr (u − L)} u≥L

= sup {(R R)(u) − λr (u)} + rL u≥L

≤ sup{(R R)(u) − λr (u)} + rL u≥0

= ((R R) λr )(0) + rL = rTmin .

This theorem provides the minimal values of playback delay Dmin and buffer Bmin , as well as the minimal look-ahead delay dmin = Tmin − Dmin for a given constant smoothing rate r < C and a given rate-latency service curve βL,C . We can also solve the dual problem, namely compute for given values of playback delay D, of the look-ahead delay d, of the playback buffer B and for a given rate-latency service curve βL,C , the minimal rate rmin which must be reserved on the network.

5.6. OPTIMAL SMOOTHING VERSUS GREEDY SHAPING

163

T HEOREM 5.5.2 (O PTIMAL CBR SMOOTHING RATE ). If σ = λr and β = βL,C with r < C, the smallest value of r, given D ≥ L, d and B ≥ (R R)(L), is R(t) (R R)(t) ∨ sup rmin = sup t+D−L t+D+d−L t>0 t>0 (R R)(t + L) − B ∨ sup . (5.22) t t>0 P ROOF : Let us first note that because of (5.19), there is no solution if D < L. On the other hand, if D ≥ L, then (5.19) implies that the rate r must be such that for all t > 0 1 D ≥ L + (R(t) − rt) r or equivalently r ≥ R(t)/(t+D−L). The latter being true for all t > 0, we must have r ≥ supt≥0 {R(t)/(t+ D − L)}. Repeating the same argument with (5.20) and (5.21), we obtain the minimal rate (5.22). In the particular case where L = 0 and r < C the network is completely transparent to the flow, and can be considered as a null network: can replace β(t) by δ0 (t). The values (5.19), (5.20) and (5.21) become, respectively, 1 (R λr )(0) r 1 ((R R) λr )(0) = r = ((R R) λr ))(0) = rTmin .

Dmin =

(5.23)

Tmin

(5.24)

Bmin

(5.25)

It is interesting to compute these values on a real video trace, such as the first trace on top of Figure 1.2.4. Since Bmin is directly proportional to Tmin because of (5.25), we show only the graphs of the values of Dmin and dmin = Tmin − Dmin , as a function of the CBR smoothing rate r on Figure 5.5. We observe three qualitative ranges of rates: (i) the very low ones where the playback delay is very large, and where look-ahead does not help in reducing it; (ii) a middle range where the playback delay can be kept quite small, thanks to the use of look-ahead and (iii) the high rates above the peak rate of the stream, which do not require any playback nor lookahead of the stream. These three regions can be found on every MPEG trace [79], and depend on the location of the large burst in the trace. If it comes sufficiently late, then the use of look-ahead can become quite useful in keeping the playback delay small.

5.6

O PTIMAL S MOOTHING VERSUS G REEDY S HAPING

An interesting question is to compare the requirements on D and B, due to the scheduling obtained in Section 5.4, which are minimal, with those that a simpler scheduling, namely the greedy shaper of Section 1.5, would create. As σ is a ‘good’ function, the solution of a greedy shaper is xshaper (t) = (σ ⊗ R)(t).

(5.26)

To be a solution for the smoothing problem, it must satisfy all constraints listed in Section 5.2. It already satisfies (5.1), (5.2) and (5.5). Enforcing (5.3) is equivalent to impose that for all t ∈ R (R β)(t − D) ≤ (σ ⊗ R)(t), which can be recast as ((R R) (β ⊗ σ))(−D) ≤ 0.

(5.27)


164 60

Dmin [sec]

50 40 30 20 10 0

0

100

200

300

400 500 rate r [kBytes/sec]

600

700

0

100

200

300

400 500 rate r [kBytes/sec]

600

700

2

dmin [sec]

1.5

1

0.5

0

Figure 5.5: Minimum playback delay Dmin and corresponding look-ahead delay dmin for a constant rate smoothing r of the MPEG-2 video trace shown on top of Figure 1.2.4. This implies that the minimal playback delay needed for a smoothing using a greedy shaping algorithm is equal to the minimal total delay Tmin , the sum of the playback and lookahead delays, for the optimal smoothing algorithm. It means that the only way an optimal smoother allows to decrease the playback delay is its ability to look ahead and send data in advance. If this look-ahead is not possible (d = 0) as for example for a live video transmission, the playback delay is the same for the greedy shaper and the optimal smoother. The last constraint that must be verified is (5.4), which is equivalent to impose that for all t ∈ R (σ ⊗ R)(t) ≤ R(t − D) + B, which can be recast as ((R ⊗ σ) R)(D) ≤ B.

(5.28)

Consequently, the minimal requirements on the playback delay and buffer using a greedy shaper instead of an optimal smoother are given by the following theorem. T HEOREM 5.6.1 (R EQUIREMENTS FOR GREEDY SHAPER ). If σ is a ‘good’ function, then the smallest values of D and B for lossless smoothing of flow R by a greedy shaper are Dshaper = Tmin = h((R R), (β ⊗ σ))

(5.29)

Bshaper = ((R ⊗ σ) R)(Dshaper ) ∈ [Bmin , σ(Dshaper )].

(5.30)

P ROOF : The expressions of Dshaper and Bshaper follow immediately from (5.27) and (5.28). The only point that remains to be shown is that Bshaper ≤ σ(Dshaper ), which we do by picking s = u in the inf below: Bshaper = (R (R ⊗ σ)) (Dshaper )

5.7. COMPARISON WITH DELAY EQUALIZATION

165

inf = sup R(s) + σ(u + Dshaper − s) − R(u) u≥0 0≤s≤u+Dshaper

≤ sup R(u) + σ(u + Dshaper − u) − R(u) u≥0

= σ(Dshaper ).

Consequently, a greedy shaper does not minimize, in general, the playback buffer requirements, although it does minimize the playback delay when look-ahead is not possible. Figure 5.6 shows the maximal solution xmax of the optimal shaper (top) and the solution xshaper of the greedy shaper (bottom) when the shaping curve is a one leaky bucket affine curve σ = γr,b , when the look-ahead delay d = 0 (no look ahead possible) and for a null network (β = δ0 ). In this case the playback delays are identical, but not the playback buffers. Another example is shown on Figure 5.7 for the MPEG-2 video trace shown on top of Figure 1.2.4. Here the solution of the optimal smoother is the minimal solution xmin . There is however one case where a greedy shaper does minimize the playback buffer: a constant rate smoothing (σ = λr ) over a null network (β = δ0 ). Indeed, in this case, (5.25) becomes Bmin = rTmin = rDshaper = σ(Dshaper ), and therefore Bshaper = Bmin . Consequently, if no look-ahead is possible and if the network is transparent to the flow, greedy shaping is an optimal CBR smoothing strategy.

5.7

C OMPARISON WITH D ELAY E QUALIZATION

A common method to implement a decoder is to first remove any delay jitter caused by the network, by delaying the arriving data in a delay equalization buffer, before using the playback buffer to compensate for fluctuations due to pre-fetching. Figure 5.8 shows such a system. If the delay equalization buffer is properly configured, its combination with the guaranteed service network results into a fixed delay network, which, from the viewpoint we take in this chapter, is equivalent to a null network. Compared to the original scenario in Figure 5.1, there are now two separate buffers for delay equalization and for compensation of prefetching. We would like to understand the impact of this separation on the minimum playback delay Dmin . The delay equalization buffer operates by delaying the first bit of data by an initial delay Deq , equal to the worst case delay through the network. We assume that the network offers a rate-latency service curve βL,C . Since the flow x is constainted by the arrival curve σ which is assumed to be a ‘good’ function, we know from Theorem 1.4.4, that the worst-case delay is Deq = h(σ, βL,C ). On the other hand, the additional part of the playback delay to compensate for fluctuations due to prefetching, denoted by Dpf , is given by (5.11) with β replaced by δ0 : Dpf = h(R, δ0 ⊗ σ) = h(R, σ). The sum of these two delays is, in general, larger than the optimal playback delay (without a separation between equalization and compensation for prefetching), Dmin , given by (5.11): Dmin = h(R, βL,C ⊗ σ).


166

σ(t) R(t) xmax(t)

b Bmin

R(t– Dmin) t

Dmin (a) Optimal smoothing solution, with playback buffer requirements

σ(t) R(t) xshaper(t)

Bshaper

R(t– Dshaper)

b

Dshaper

t

(a) Greedy shaper solution, with playback buffer requirements

Figure 5.6: In bold, the maximal solution (top figure) and minimal solution (bottom figure) to the smoothing problem with a null network, no look-ahead and an affine smoothing curve σ = γr,b . Consider the example of Figure 5.9, where σ = γr,b with r < C. Then one easily computes the three delays Dmin , Deq and Dpf , knowing that βL,C ⊗ σ = δL ⊗ λC ⊗ γr,b = δL ⊗ (λC ∧ γr,b ) = (δL ⊗ λC ) ∧ (δL ⊗ γr,b ) = βL,C ∧ (δL ⊗ γr,b ). One clearly has Dmin < Deq + Dpf : separate delay equalization gives indeed a larger overall playback delay. In fact, looking carefully at the figure (or working out the computations), we can observe that the combination of delay equalization and compensation for prefetching in a single buffer accounts for the busrtiness of the (optimally) smoothed flow only once. This is another instance of the “pay bursts only

5.7. COMPARISON WITH DELAY EQUALIZATION

167

6

5

x 10

Video trace R(t) Optimal shaping

4.5

4

Cummulative # of bits

3.5

3

2.5

2

1.5

1

0.5

0

0

50

100

150

200

250 300 Frame number

350

400

450

500

6

5

x 10

Video trace R(t) Optimal smoothing

4.5

4

Cummulative # of bits

3.5

3

2.5

2

1.5

1

0.5

0

0

50

100

150

200 250 Frame number

300

350

400

450

Figure 5.7: Example of optimal shaping versus optimal smoothing for the MPEG-2 video trace shown on top of Figure 1.2.4. The example is for a null network and a smoothing curve σ = γP,M ∧γr,b with M = 416 bytes, P = 600 kBytes/sec, r = 300 kBytes/sec and b = 80 kBytes. The figure shows the optimal shaper [resp. smoother] output and the original signal (video trace), shifted by the required playback delay. The playback delay is 2.76 sec for optimal shaping (top) and 1.92 sec for optimal smoothing (bottom). Guaranteed service network Video Server


Network σ β

x(t) R(t+d) Smoother

R(t-D) Delay y(t) Client equalizer playback buffer

Figure 5.8: Delay equalization at the receiver.


168

once” phenomenon, which we have already met in Section 1.4.3.

β(t) D min R(t)

σ(t) (σ⊗β)(t)

D pf

Deq t

Figure 5.9: Delays Dmin , Deq and Dpf for a rate-latency service curve βL,C and an affine smoothing curve σ = γr,b . We must however make – once again – an exception for a constant rate smoothing. Indeed, if σ = λr (with r < C), then Dpf is given by (5.23) and Dmin by (5.19), so that

and therefore Dmin = Deq optimal playback delay.

5.8

Deq = h(λr , βL,C ) = L 1 (R λr )(0) Dpf = r 1 Dmin = L + (R λr )(0) r + Dpf . In the CBR case, separate delay equalization is thus able to attain the

L OSSLESS S MOOTHING OVER T WO N ETWORKS

We now consider the more complex setting where two networks separate the video server from the client: the first one is a backbone network, offering a service curve β1 to the flow, and the second one is a local access network, offering a service curve β2 to the flow, as shown on Figure 5.10. This scenario models intelligent, dynamic caching often done at local network head-ends. We will compute the requirements on D, d, B and on the buffer X of this intermediate node in Subsection 5.8.1. Moreover, we will see in Subsection 5.8.2 that for constant rate shaping curves and rate-latency service curves, the size of the client buffer B can be reduced by implementing a particular smoothing strategy instead of FIFO scheduling at the intermediate node. Two flows need therefore to be computed: the first one x1 (t) at the input of the backbone network, and the second one x2 (t) at the input of the local access network, as shown on Figure 5.10. The constraints on both flows are now as follows:

5.8. LOSSLESS SMOOTHING OVER TWO NETWORKS

Video Server

Backbone Network σ1

R(t+d)

Intermediate Access storage Network Node σ2

β1

Smoother

x1(t)

169

X y1(t)


β2

x2(t) Smoother

y2(t)

B R(t-D)

Client playback buffer

Figure 5.10: Smoothing over two networks with a local caching node. • Causal flow x1 : This constraint is the same as (5.1), but with x replaced by x1 : x1 (t) ≤ δ0 (t),

(5.31)

• Smoothness constraint: Both flows x1 and x2 are constrained by two arrival curves σ1 and σ2 : x1 (t) ≤ (x1 ⊗ σ1 )(t)

(5.32)

x2 (t) ≤ (x2 ⊗ σ2 )(t).

(5.33)

• No playback and intermediate server buffers underflow: The data is read out from the playback buffer after D unit of times at a rate given by R(t − D), which implies that y2 (t) ≥ R(t − D). On the other hand, the data is retrieved from the intermediate server at a rate given by x2 (t), which implies that y1 (t) ≥ x2 (t). As we do not know the expressions of the outputs of each network, but only a service curve β1 and β2 for each of them, we can replace y1 by x1 ⊗ β1 and y2 by x2 ⊗ β2 , and reformulate these two constraints by x2 (t) ≤ (x1 ⊗ β1 )(t)

(5.34)

x2 (t) ≥ (R β2 )(t − D).

(5.35)

• No playback and intermediate server buffers overflow: The size of the playback and cache buffers are limited to B and X, respectively, and to prevent any overflow of the buffer, we must impose that y1 (t) − x2 (t) ≤ X and y2 (t) − R(t − D) ≤ B for all t ≥ 0. Again, we do not know the exact value of y1 and y2 , but we know that they are bounded by x1 and x2 , respectively, so that the constraints becomes, for all t ≥ 0, x1 (t) ≤ x2 (t) + X

(5.36)

x2 (t) ≤ R(t − D) + B.

(5.37)

• Look-ahead delay constraint: this constraint is the same as in the single network case: x1 (t) ≤ R(t + d).

5.8.1

(5.38)

M INIMAL R EQUIREMENTS ON THE D ELAYS AND B UFFER S IZES FOR T WO N ETWORKS

Inequalities (5.31) to (5.38) can be recast as three sets of inequalities as follows: x1 (t) ≤ δ0 (t) ∧ R(t + d) ∧ (σ1 ⊗ x1 )(t) ∧ (x2 (t) + X)

(5.39)

x2 (t) ≤ {R(t − D) + B} ∧ (β1 ⊗ x1 )(t) ∧ (σ2 ⊗ x2 )(t)

(5.40)

x2 (t) ≥ (R β2 )(t − D).

(5.41)


170

We use the same technique for solving this problem sa in Section 5.3, except that now the dimension of the system J is 2 instead of 1. With T denoting transposition, let us introduce the following notations: x(t) = [x1 (t) x2 (t)]T a(t) = [δ0 (t) ∧ R(t + d) R(t − D) + B]T b(t) = [0 (R β2 )(t − D)]T Σ(t) =

σ1 (t) δ0 (t) + X σ2 (t) β1 (t)

.

With these notations, the set of inequalities (5.39), (5.40) and (5.41) can therefore be recast as x ≤ a ∧ (Σ ⊗ x) x ≥ b.

(5.42) (5.43)

We will follow the same approach as in Section 5.3: we first compute the maximal solution of (5.42) and then derive the constraints on D, T (and hence d), X and B ensuring the existence of this solution. We apply thus Theorem 4.3.1 again, but this time in the two-dimensional case, to obtain an explicit formulation of the maximal solution of (5.42). We get xmax = C Σ (a) = (Σ ⊗ a)

(5.44)

where Σ is the sub-additive closure of Σ, which is, as we know from Section 4.2, Σ = inf {Σ(n) }

(5.45)

n∈N

where Σ(0) = D0 and Σ(n) denotes the nth self-convolution of Σ. Application of (5.45) to matrix Σ is straightforward, but involves a few manipulations which are skipped. Denoting

(n+1) α = σ1 ⊗ σ2 ⊗ inf β1 + nX (5.46) n∈N

= σ1 ⊗ σ2 ⊗ β1 ⊗ β1 + X, we find that

σ1 ∧(α+ X) (σ1 ⊗σ2 +X)∧(α + 2X) Σ= α σ2 ∧ (α + X)

and therefore the two coordinates of the maximal solution of (5.42) are x1 max (t) = σ1 (t) ∧ {α(t) + X} ∧ (σ1 ⊗ R)(t + d) ∧ {(α ⊗ R)(t + d) + X} ∧ {(σ1 ⊗ σ2 ⊗ R)(t − D) + B + X} ∧ {(α ⊗ R)(t − D) + B + 2X}

(5.47)

x2 max (t) = α(t) ∧ (α ⊗ R)(t + d) ∧ {(σ2 ⊗ R)(t − D) + B} ∧ {(α ⊗ R)(t − D) + B + X} .

(5.48)

Let us mention that an alternative (and computationally simpler) approach to obtain (5.47) and (5.48) would have been to first compte the maximal solution of (5.40), as a function of x1 , and next to replace x2 in (5.39) by this latter value. We can now express the constraints on X, B, D and d that will ensure that a solution exists by requiring that (5.48) be larger than (5.41). The result is stated in the following theorem, whose proof is similar to that of Theorem 5.3.1.

5.8. LOSSLESS SMOOTHING OVER TWO NETWORKS

171

T HEOREM 5.8.1. The lossless smoothing of a flow to (sub-additive) curves σ1 and σ2 , respectively, over two networks offering service curves β1 and β2 has a solution if and only if the D, T , X and B verify the following set of inequalities, with α defined by (5.46): (R (α ⊗ β2 )(−D) ≤ 0

(5.49)

((R R) (α ⊗ β2 )) (−T ) ≤ 0

(5.50)

((R R) (σ2 ⊗ β2 )) (0) ≤ B

(5.51)

((R R) (α ⊗ β2 )) (0) ≤ B + X.

5.8.2

(5.52)

O PTIMAL C ONSTANT R ATE S MOOTHING OVER T WO N ETWORKS

Let us compute the values of Theorem 5.8.1 in the case of two constant rate (CBR) smoothing curves σ1 = λr1 and σ2 = λr2 . We assume that each network offers a rate-latency service curve βi = βLi ,Ci , i = 1, 2. We assume that ri ≤ Ci In this case the optimal values of D, T and B become the following ones, depending on the value of X. T HEOREM 5.8.2. Let r = r1 ∧ r2 . Then we have the following three cases depending on X: (i) If X ≥ rL1 , then Dmin , Tmin and Bmin are given by 1 Dmin = L1 + L2 + (R λr )(0) r 1 Tmin = L1 + L2 + ((R R) λr )(0) r Bmin = ((R R) λr2 )(L2 ) ∨ {((R R) λr )(L1 + L2 ) − X} ≤ ((R R) λr )(L2 ).

(5.53) (5.54) (5.55)

(ii) If 0 < X < rL1 then Dmin , Tmin and Bmin are bounded by L1 X + L2 + (R λ X )(0) ≤ Dmin L1 r X L1 (R λ X )(0) ≤ L1 + L2 + L1 X L1 X + L2 + ((R R) λ X )(0) ≤ Tmin L1 r X L1 ((R R) λ X )(0) ≤ L1 + L2 + L1 X ((R R) λ X )(L1 + L2 ) − r2 L1 ≤ Bmin

(5.56)

(5.57)

L1

≤ ((R R) λ X )(L2 )

(5.58)

L1

(iii) Let K be duration of the stream. If X = 0 < rL1 then Dmin = K. (n+1)

(n+1)

Proof. One easily verifies that δL1 = δ(n+1)L1 and that λC1 = λC1 . Since β1 = βL1 ,C1 = δL1 ⊗ λC1 , and since r = r1 ∧ r2 ≤ C1 , (5.46) becomes α = λr ⊗ inf δ(n+1)L1 ⊗ λC1 + nX n∈N

= δL1 ⊗ inf {δnL1 ⊗ λr + nX} . n∈N

(i) If X ≥ rL1 , then for t ≥ nL1 (δnL1 ⊗ λr )(t) + nX = λr (t − nL1 ) + nX = rt + n(X − rL1 ) ≥ rt = λr (t)

(5.59)


172 whereas for 0 ≤ t < nL1

(δnL1 ⊗ λr )(t) + nX = λr (t − nL1 ) + nX = nX ≥ nrL1 > rt = λr (t). Consequently, for all t ≥ 0, α(t) ≥ (δL1 ⊗ λr )(t). On the other hand, taking n = 0 in the infimum in (5.59) yields that α ≤ δL1 ⊗ λr . Combining these two inequalities, we get that α = δL1 ⊗ λr and hence that α ⊗ β2 = δL1 ⊗ λr ⊗ δL2 ⊗ λr2 = δL1 +L2 ⊗ λr = βL1 +L2 ,r .

(5.60)

Inserting this last relation in (5.49) to (5.52), and using Lemma 5.5.1 we establish (5.53), (5.54) and the equality in (5.55). The inequality in (5.55) is obtained by noticing that r2 ≥ r and that ((R R) λr )(L1 + L2 ) − X = sup{(R R)(u + L1 + L2 ) − ru} − X u≥0

=

sup {(R R)(v + L2 ) − r(v − L1 )} − X

v≥L1

≤ sup{(R R)(v + L2 ) − rv} + (rL1 − X) v≥0

≤ ((R R) λr )(L2 ). (ii) If 0 < X < rL1 , the computation of α does not provide a rate-latency curve anymore, but a function that can be bounded below and above by the two following rate-latency curves: βL1 ,X/L1 ≤ α ≤ βX/r,X/L1 . Therefore, replacing (5.60) by δL1 +L2 ⊗ λ X ≤ α ⊗ β2 ≤ δ X +L2 ⊗ λ X , L1

r

L1

and applying Lemma 5.5.1 to both bounding rate-latency curves βL1 ,X/L1 and βX/r,X/L1 , we get respectively the lower and upper bounds (5.56) to (5.58). (iii) If X = 0 and rL1 > 0 then (5.59) yields that α(t) = 0 for all t ≥ 0. In this case (5.49) becomes supu≥0 {R(u − D)} ≤ 0. This is possible only if D is equal to the duration of the stream. It is interesting to examine these results for two particular values of X. The first one is X = ∞. If the intermediate server is a greedy shaper whose output is x2 (t) = (σ2 ⊗ y1 )(t), one could have applied Theorem 5.5.1 with σ2 = λr and β = β1 ⊗ σ2 ⊗ β2 = δL1 +L2 ⊗ λr2 = βL1 +L2 ,r2 to find out that D and T are still given by (5.53) and (5.54) but that B = ((R R) λr )(L1 + L2 ) is larger than (5.55). Using the caching scheduling (5.48) instead of a greedy shaping one allows therefore to decrease the playback buffer size, but not the delays. The buffer X of the intermediate node does not need to be infinite, but can be limited to rL1 . The second one is X = 0. Then whatever the rate r > 0, if L1 > 0, the playback delay is the length of the stream, which makes streaming impossible in practice. When L1 = L2 = 0 however (in which case we have two null networks) X = rL1 = 0 is the optimal intermediate node buffer allocation. This was shown in [69](Lemma 5.3) using another approach. We see that when L1 > 0, this is no longer the case.

5.9


The first application of network calculus to optimal smoohting is found in [54], for an unlimited value of the look-ahead delay. The minimal solution (5.17) is shown to be an optimal smoothing scheme. The


173

computation of the minimum look-ahead delay, and of the maximal solution, is done in [79]. Network calculus allows to retrieve some results found using other methods, such as the optimal buffer allocation of the intermdiate node for two null networks computed in [69]. It also allows to extend these results, by computing the full set of optimal schedules and by taking into account non null networks, as well as by using more complex shaping curves σ than constant rate service curves. For example, with the Resource Reservation Protocol (RSVP), σ is derived from the T-SPEC field in messages used for setting up the reservation, and is given by σ = γP,M ∧ γr,b , where M is the maximum packet size, P the peak rate, r the sustainable rate and b the burst tolerance, as we have seen in Section 1.4.3. The optimal T-SPEC field is computed in [54]. More precisely, the following problem is solved. As assumed by the Intserv model, every node offers a service of the form βL,C for some latency L and rate C, with the latency parameter L depending on the rate C according to L = Cρ0 + D0 . The constants C0 and D0 depends on the route taken by the flow throughout the network. Destinations choose a target admissible network delay Dnet . The choice of a specific service curve βL,C (or equivalently, of a rate parameter C) is done during the reservation phase and cannot be known exactly in advance. The algorithm developed in [54] computes the admissible choices of σ = γP,M ∧ γr,b and of Dnet in order to guarantee that the reservation that will subsequently be performed ensures a playback delay not exceeding a given value D.

174


C HAPTER 6

AGGREGATE S CHEDULING

6.1

I NTRODUCTION

Aggregate scheduling arises naturally in many case. Let us just mention here the differentiated services framework (Section 2.4 on Page 86) and high speed switches with optical switching matrix and FIFO outputs. The state of the art for aggregate multiplexing is not very rich. In this chapter, we give a panorama of results, a number of which are new. In a first step (Section 6.2), we evaluate how an arrival curve is transformed through aggregate multiplexing; we give a catalog of results, when the multiplexing node is either a service curve element with FIFO scheduling, or a Guaranteed Rate node (Section 2.1.3), or a service curve element with strict service curve property. This provides many simple, explicit bounds which can be used in practice. In a second step (Section 6.3), we consider a global network using aggregate multiplexing (see assumptions below); given constraints at the inputs of the network, can we obtain some bounds for backlog and delay ? Here, the story is complex. The question of delay bounds for a network with aggregate scheduling was first raised by Chang [8]. For a given family of networks, we call critical load factor νcri a value of utilization factor below which finite bounds exist, and above which there exist unstable networks, i.e., networks whose backlog grow to infinity. For feed-forward networks with aggregate multiplexing, an iterative application of Section 6.2 easily shows that νcri = 1. However, many networks are not feed-forward, and this result does not hold in general. Indeed, and maybe contrary to intuition, Andrews [3] gave some examples of FIFO networks with νcri < 1. Still, the iterative application of Section 6.2, augmented with a time-stopping argument, provides lower bounds of νcri (which are less than 1). In a third step (Section 6.4), we give a number of cases where we can say more. We recall the result in Theorem 2.4.1 on Page 88, which says that, for a general network with either FIFO service curve elements, 1 where h is abound on the number of hops seen by any flow. Then, or with GR nodes, we have νcri ≥ h−1 in Section 6.4.1, we show that the unidirectional ring always always has νcri = 1; thus, and this may be considered a surprise, the ring is not representative of non feed-forward topologies. This result is actually true under the very general assumption that the nodes on the ring are service curve elements, with any values of link speeds, and with any scheduling policy (even non FIFO) that satisfies a service curve property. As far as we know, we do not really understand why the ring is always stable, and why other topologies may not be. Last, and not least surprising, we present in Section 6.4.2 a particular case, originally found by Chlamtac, Faragó, Zhang, and Fumagalli [15], and refined by Zhang [83] and Le Boudec and Hébuterne [52] which shows that, for a homogeneous network of FIFO nodes with constant size packets, strong rate limitations at 175

CHAPTER 6. AGGREGATE SCHEDULING

176

all sources have the effect of providing simple, closed form bounds.

6.2

T RANSFORMATION OF A RRIVAL C URVE THROUGH AGGREGATE S CHEDUL ING

Consider a number of flows served as an aggregate in a common node. Without loss of generality, we consider only the case of two flows. Within an aggregate, packets are served according to some unspecified arbitration policy. In the following sub-sections, we consider three additional assumptions.

6.2.1

AGGREGATE M ULTIPLEXING IN A S TRICT S ERVICE C URVE E LEMENT

The strict service curve property is defined in Definition 1.3.2 on Page 21. It applies to some isolated schedulers, but not to complex nodes with delay elements. T HEOREM 6.2.1 (B LIND MULTIPLEXING ). Consider a node serving two flows, 1 and 2, with some unknown arbitration between the two flows. Assume that the node guarantees a strict service curve β to the aggregate of the two flows. Assume that flow 2 is α2 -smooth. Define β1 (t) := [β(t) − α2 (t)]+ . If β1 is wide-sense increasing, then it is a service curve for flow 1. P ROOF :

The proof is a straightforward extension of that of Proposition 1.3.4 on Page 21.

We have seen an example in Section 1.3.2: if β(t) = Ct (constant rate server or GPS node) and α2 = γr,b (constraint by one leaky bucket) then the service curve for flow 1 is the rate-latency service curve with rate b C − r and latency C−r . Note that the bound in Theorem 6.2.1 is actually for a preemptive priority scheduler where flow 1 has low priority. It turns out that if we have no other information about the system, it is the only bound we can find. For completeness, we give the following case. C OROLLARY 6.2.1 (N ON PREEMPTIVE PRIORITY NODE ). Consider a node serving two flows, H and L, with non-preemptive priority given to flow H. Assume that the node guarantees a strict service curve β to the L ]+ aggregate of the two flows. Then the high priority flow is guaranteed a service curve βH (t) = [β(t)−lmax L where lmax is the maximum packet size for the low priority flow. If in addition the high priority flow is αH -smooth, then define βL by βL (t) = [β(t) − αH (t)]+ . If βL is wide-sense increasing, then it is a service curve for the low priority flow. P ROOF : The first part is an immediate consequence of Theorem 6.2.1. The second part is proven in the same way as Proposition 1.3.4. If the arrival curves are affine, then the following corollary of Theorem 6.2.1 expresses the burstiness increase due to multiplexing. C OROLLARY 6.2.2 (B URSTINESS I NCREASE DUE TO B LIND M ULTIPLEXING ). Consider a node serving two flows in an aggregate manner. Assume the aggregate is guaranteed a strict service curve βR,T . Assume also that flow i is constrained by one leaky bucket with parameters (ρi , σi ). If ρ1 + ρ2 ≤ R the output of the first flow is constrained by a leaky bucket with parameters (ρ1 , b∗1 ) with b∗1 = σ1 + ρ1 T + ρ1

σ 2 + ρ2 T R − ρ2

Note that the burstiness increase contains a term ρ1 T that is found even if there is no multiplexing; the 2 +ρ2 T second term ρ1 σR−ρ comes from multiplexing with flow 2. Note also that if we further assume that the 2 node is FIFO, then we have a better bound (Section 6.2.2).

6.2. TRANSFORMATION OF ARRIVAL CURVE THROUGH AGGREGATE SCHEDULING

177

P ROOF : From Theorem 6.2.1, the first flow is guaranteed a service curve βR ,T with R = R − ρ2 and 2 +T ρ2 . The result follows from a direct application of Theorem 1.4.3 on Page 23. T = σR−ρ 2 D O WE NEED THAT THE SERVICE CURVE PROPERTY BE STRICT ? If we relax the assumption that the service curve property is strict, then the above results do not hold. A counter-example can be built as follows. All packets have the same size, 1 data unit, and input flows have a peak rate equal to 1. Flow 1 sends one packet at time 0, and then stops. The node delays this packet forever. With an obvious notation, we have, for t ≥ 0: R1 (t) = min(t, 1) and R1 (t) = 0 Flow 2 sends one packet every time unit, starting at time t = 1. The output is a continuous stream of packets, one per time unit, starting from time 1. Thus R2 (t) = (t − 1)+ and R2 (t) = R2 (t) The aggregate flows are, for t ≥ 0: R(t) = t and R (t) = (t − 1)t In other words, the node offers to the aggregate flow a service curve δ1 . Obviously, Theorem 6.2.1 does not apply to flow 1: if it would, flow 1 would receive a service curve (δ1 − λ1 )+ = δ1 , which is not true since it receives 0 service. We can interpret this example in the light of Section 1.4.4 on Page 29: if the service curve property would be strict, then we could bound the duration of the busy period, which would give a minimum service guarantee to low priority traffic. We do not have such a bound on this example. In Section 6.2.2 we see that if we assume FIFO scheduling, then we do have a service curve guarantee.

6.2.2

AGGREGATE M ULTIPLEXING IN A FIFO S ERVICE C URVE E LEMENT

Now we relax the strict service curve property; we assume that the node guarantees to the aggregate flow a minimum service curve, and in addition assume that it handles packets in order of arrival at the node. We find some explicit closed forms bounds for some simple cases. P ROPOSITION 6.2.1 (FIFO M INIMUM S ERVICE C URVES [20]). Consider a lossless node serving two flows, 1 and 2, in FIFO order. Assume that packet arrivals are instantaneous. Assume that the node guarantees a minimum service curve β to the aggregate of the two flows. Assume that flow 2 is α2 -smooth. Define the family of functions βθ1 by βθ1 (t) = [β(t) − α2 (t − θ)]+ 1{t>θ} Call R1 (t), R1 (t) the input and output for flow 1. Then for any θ ≥ 0 R1 ≥ R1 ⊗ βθ1

(6.1)

If βθ1 is wide-sense increasing, flow 1 is guaranteed the service curve βθ1 The assumption that packet arrivals are instantaneous means that we are either in a fluid system (one packet is one bit or one cell), or that the input to the node is packetized prior to being handled in FIFO order.


178

P ROOF : We give the proof for continuous time and assume that flow functions are left-continuous. All we need to show is Equation (6.1). Call Ri the flow i input, R = R1 + R2 , and similarly Ri , R the output flows. Fix some arbitrary parameter θ and time t. Define u := sup{v : R(v) ≤ R (t)} Note that u ≤ t and that

R(u) ≤ R (t) and R(u+ ) ≥ R (t)

(6.2)

where Rr (u) = inf v>u [R(v)] is the limit to the right of R at u. (Case 1) consider the case where u = t. It follows from the above and from R ≤ R that R1 (t) = R1 (t). Thus for any θ, we have R1 (t) = R1 (t) + βθ1 (0) which shows that R1 (t) ≥ (R1 ⊗ βθ1 )(t) in that case. (Case 2), assume now that u < t. We claim that R1 (u) ≤ R1 (t)

(6.3)

Indeed, if this is not true, namely, R1 (u) > R1 (t), it follows from the first part of Equation (6.2) that R2 (u) < R2 (t). Thus some bits from flow 2 arrived after time u and departed by time t, whereas all bits of flow 1 arrived up to time u have not yet departed at time t. This contradicts our assumption that the node is FIFO and that packets arrive instantaneously. Similarly, we claim that

(R2 )r (u) ≥ R2 (t)

Indeed, otherwise x := R2 (t) − (R2 )r (u) > 0 and there is we have R2 (v) < R2 (t) − x2 . From Equation (6.2), we can then R1 (v) + R2 (v) ≥ R (t) − x4 . It follows that

(6.4)

some v0 ∈ (u, t] such that for any v ∈ (u, v0 ] find some v1 ∈ (u, v0 ] such that if v ∈ (u, v1 ]

R1 (v) ≥ R1 (t) +

x 4

Thus we can find some v with R1 (v) > R1 (t) whereas R2 (v) < R2 (t), which contradicts the FIFO assumption. Call s a time such that R (t) ≥ R(s) + β(t − s). We have R(s) ≤ R (t) thus s ≤ u. (Case 2a) Assume that u < t − θ thus also t − s > θ. From Equation (6.4) we derive R1 (t) ≥ R1 (s) + β(t − s) + R2 (s) − R2 (t) ≥ R1 (s) + β(t − s) + R2 (s) − (R2 )r (u) Now there exist some > 0 such that u + ≤ t − θ, thus (R2 )r (u) ≤ R2 (t − θ) and R1 (t) ≥ R1 (s) + β(t − s) − α2 (t − s − θ) It follows from Equation (6.3) that which shows that

R1 (t) ≥ R1 (s) R1 (t) ≥ R1 (s) + βθ1 (t − s)

(Case 2b) Assume that u ≥ t − θ. By Equation (6.3): R1 (t) ≥ R1 (u) = R1 (u) + βθ1 (t − u) We cannot conclude from Proposition 6.2.1 that inf θ βθ1 is a service curve. However, we can conclude something for the output.

6.2. TRANSFORMATION OF ARRIVAL CURVE THROUGH AGGREGATE SCHEDULING

179

P ROPOSITION 6.2.2 (B OUND FOR O UTPUT WITH FIFO). Consider a lossless node serving two flows, 1 and 2, in FIFO order. Assume that packet arrivals are instantaneous. Assume that the node guarantees to the aggregate of the two flows a minimum service curve β. Assume that flow 2 is α2 -smooth. Define the family of functions as in Proposition 6.2.1. Then the output of flow 1 is α1∗ -smooth, with & ' α1∗ (t) = inf α1 βθ1 (t) θ≥0

P ROOF : Observe first that the network calculus output bound holds even if β is not wide-sense increasing. Thus, from Proposition 6.2.1, we can conclude that α1 βθ1 is an arrival curve for the output of flow 1. This is true for any θ. We can apply the last proposition and obtain the following practical result. T HEOREM 6.2.2 (B URSTINESS I NCREASE DUE TO FIFO, G ENERAL C ASE ). Consider a node serving two flows, 1 and 2, in FIFO order. Assume that flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness σ1 , and flow 2 is constrained by a sub-additive arrival curve α2 . Assume that the node guarantees to the aggregate of the two flows a rate latency service curve βR,T . Call ρ2 := inf t>0 1t α2 (t) the maximum sustainable rate for flow 2. If ρ1 + ρ2 < R, then at the output, flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness b∗1 with ˆ B b∗1 = σ1 + ρ1 T + R and ˆ = sup [α2 (t) + ρ1 t − Rt] B t≥0

The bound is a worst case bound. P ROOF : (Step 1) Define βθ1 as in Proposition 6.2.1. Define B2 = supt≥0 [α2 (t) − Rt]. Thus B2 is the buffer that would be required if the latency T would be 0. We first show the following if θ ≥

B2 + T then for t ≥ θ : βθ1 (t) = Rt − RT − α2 (t − θ) R

(6.5)

To prove this, call φ(t) the right hand-side in Equation (6.5), namely, for t ≥ θ define φ(t) = Rt − α2 (t − θ) − RT . We have inf φ(t) = inf [Rv − α2 (v) − RT + Rθ] v>0

t>θ

From the definition of B2 :

inf φ(t) = −B2 + Rθ − RT

t>θ

If θ ≥

B2 R

+ T then φ(t) ≥ 0 for all t > θ. The rest follows from the definition of βθ1 .

(Step 2) We apply the second part of Proposition 6.2.1 with θ = flow 1 is given by α1∗ = λρ1 ,σ1 βθ1

ˆ B R

+ T . An arrival curve for the output of

ˆ ≤ B2 , and therefore β 1 (t) = Rt − RT − α2 (t − θ) for We now compute α1∗ . First note that obviously B θ t ≥ θ. α1∗ is thus defined for t > 0 by α1∗ (t) = sup ρ1 t + σ1 + ρ1 s − βθ1 (s) = ρ1 t + σ1 + sup ρ1 s − βθ1 (s) s≥0

s≥0


180 Define ψ(s) := ρ1 s − βθ1 (s). Obviously:

sup [ψ(s)] = ρ1 θ

s∈[0,θ]

Now from Step 1, we have sup[ψ(s)] = sup [ρ1 s − Rs + RT + α2 (s − θ)] s>θ

s>θ

= sup [ρ1 v − Rvα2 (v)] + (ρ1 − R)θ + RT v>0

ˆ the former is equal to From the definition of B, ˆ + (ρ1 − R)θ + RT = ρ1 θ sup[ψ(s)] = B s>θ

which shows the burstiness bound in the theorem. ˆ − (R − ρ1 )θ. ˆ Define ˆ = (α2 )r (θ) (Step 3) We show that the bound is attained. There is a time a θˆ such that B flow 2 to be greedy up to time θˆ and stop from there on:

R2 (t) = α2 (t) for t ≤ θˆ ˆ for t > θˆ R2 (t) = (R2 )r (θ)

Flow 2 is α2 -smooth because α2 is sub-additive. Define flow 1 by

R1 (t) = ρ1 t for t ≤ θˆ R1 (t) = ρ1 t + σ1 for t > θˆ

Flow 1 is λρ1 ,σ1 -smooth as required. Assume the server delays all bits by T at time 0, then after time T operates with a constant rate R, until time θˆ + θ, when it becomes infinitely fast. Thus the server satisfies ˆ + RT . Thus all flow-2 the required service curve property. The backlog just after time θˆ is precisely B ˆ B bits that arrive just after time θˆ are delayed by R + T = θ. The output for flow 1 during the time interval ˆ θˆ + t], thus there are ρ1 t + b∗ such bits, for any (θˆ + θ, θˆ + θ + t] is made of the bits that have arrived in (θ, 1

t. The following corollary is an immediate consequence. C OROLLARY 6.2.3 (B URSTINESS I NCREASE DUE TO FIFO). Consider a node serving two flows, 1 and 2, in FIFO order. Assume that flow i is constrained by one leaky bucket with rate ρi and burstiness σi . Assume that the node guarantees to the aggregate of the two flows a rate latency service curve βR,T . If ρ1 + ρ2 < R, then flow 1 has a service curve equal to the rate latency function with rate R − ρ2 and latency T + σR2 and at the output, flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness b∗1 with σ2 b∗1 = σ1 + ρ1 T + R

Note that this bound is better than the one we used in Corollary 6.2.2 (but the assumptions are slightly different). Indeed, in that case, we would obtain the rate-latency service curve with the same rate R − ρ2 2 +ρ2 T but with a larger latency: T + σR−ρ instead of T + σR2 . The gain is due to the FIFO assumption. 2

6.3. STABILITY AND BOUNDS FOR A NETWORK WITH AGGREGATE SCHEDULING

6.2.3

181

AGGREGATE M ULTIPLEXING IN A GR N ODE

We assume now that the node is of the Guaranteed Rate type (Section 2.1.3 on Page 70). A FIFO service curve element with rate-latency service curve satisfies this assumption, but the converse is not true (Theorem 2.1.3 on Page 71). T HEOREM 6.2.3. Consider a node serving two flows, 1 and 2 in some aggregate manner. Arbitration between flows is unspecified, but the node serves the aggregrate as a GR node with rate R and latency T . Assume that flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness σ1 , and flow 2 is constrained by a sub-additive arrival curve α2 . Call ρ2 := inf t>0 1t α2 (t) the maximum sustainable rate for flow 2. If ρ1 + ρ2 < R, then at the output, flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness b∗1 with ˆ b∗ = σ1 + ρ1 T + D 1

and

ˆ = sup[ α2 (t) + ρ1 t + σ1 − t] D R t>0

ˆ + T . Thus an arrival P ROOF : From Theorem 2.1.4 on Page 71, the delay for any packet is bounded by D ˆ curve at the output of flow 1 is α1 (t + D). C OROLLARY 6.2.4. Consider a node serving two flows, 1 and 2 in some aggregate manner. Arbitration between flows is unspecified, but the node serves the aggregrate as a GR node with rate R and latency T . Assume that flow i is constrained by one leaky bucket with rate ρi and burstiness σi . If ρ1 + ρ2 < R, then, at the output, flow 1 is constrained by one leaky bucket with rate ρ1 and burstiness b∗1 with

σ1 + σ2 ∗ b1 = σ1 + ρ1 T + R We see that the bound in this section is less good than Corollary 6.2.3 (but the assumptions are more general).

6.3

S TABILITY AND B OUNDS FOR A N ETWORK WITH AGGREGATE S CHEDUL ING

6.3.1

T HE I SSUE OF S TABILITY

In this section we consider the following global problem: Given a network with aggregate scheduling and arrival curve constraints at the input (as defined in the introduction) can we find good bounds for delay and backlog ? Alternatively, when is a network with aggregate scheduling stable (i.e., the backlog remains bounded) ? As it turns out today, this problem is open in many cases. In the rest of the chapter, we make the following assumptions. A SSUMPTION AND N OTATION • Consider a network with a fixed number I of flows, following fixed paths. The collection of paths is called the topology of the network. A network node is modeled as a collection of output buffers, with no contention other than at the output buffers. Every buffer is associated with one unidirectional link that it feeds.


182

• Flow i is constrained by one leaky bucket of rate ρi and burstiness σi at the input. • Inside the network, flows are treated as an aggregate by the network; within an aggregate, packets are served according to some unspecified arbitration policy. We assume that the node is such that the aggregate of all flows receives a service curve at node m equal to the rate-latency function with rate rm and latency em . This does not imply that the node is work-conserving. Also note that we do not require, unless otherwise specified, that the service curve property be strict. In some parts of the chapter, we make additional assumptions, as explained later. em accounts for the latency on the link that exits node m; it also account for delays due to the scheduler at node m. • We write i m to express that node m is on the route of flow i. For any node m, define ρ(m) = ρ(m) ρ(m) im ρi . The utilization factor of link m is rm and the load factor of the network is ν = maxm rm . • The bit rate of the link feeding node m is Cm < +∞, with Cm ≥ rm . In the context of the following definition, we call “network” N a system satisfying the assumptions above, where all parameters except ρi , σi , rm , em are fixed. In some cases (Section 6.3.2), we may add additional constraints on these parameters. D EFINITION 6.3.1 (C RITICAL L OAD FACTOR ). We say that νcri is the critical load factor for a network N if • for all values of ρi , σi , rm , em such that ν < νcri , N is stable • there exists values of ρi , σi , rm , em with ν > νcri such that N is unstable. It can easily be checked that νcri is unique for a given network N . It is also easy to see that for all well defined networks, the critical load factor is ≤ 1. However, Andrews gave in [3] an example of a FIFO network with νcri < 1. The problem of finding the critical load factor, even for the simple case of a FIFO network of constant rate servers, seems to remain open. Hajek [37] shows that, in this last case, the problem can be reduced to that where every source i sends a burst σi instantly at time 0, then sends at a rate limited by ρi . In the rest of this section and in Section 6.4, we give lower bounds on νcri for some well defined sub-classes. F EED -F ORWARD N ETWORKS A feed-forward network is one in which the graph of unidirectional links has no cycle. Examples are interconnection networks used inside routers or multiprocessor machines. For a feed-forward network made of strict service curve element or GR nodes, νcri = 1. This derives from applying the burstiness increase bounds given in Section 6.2 repeatedly, starting from network access points. Indeed, since there is no loop in the topology, the process stops and all input flows have finite burstiness. A L OWER B OUND ON THE C RITICAL L OAD FACTOR It follows immediately from Theorem 2.4.1 on 1 , where h is Page 88 that for a network of GR nodes (or FIFO service curve elements), we have νcri ≥ h−1 the maximum hop count for any flow. A slightly better bound can be found if we exploit the values of the peak rates Cm (Theorem 2.4.2).

6.3.2

T HE T IME S TOPPING M ETHOD

For a non feed-forward network made of strict service curve element or GR nodes, we can find a lower bound on νcri (together with bounds on backlog or delay), using the time stopping method. It was introduced by Cruz in [22] together with bounds on backlog or delay. We illustrate the method on a specific example, shown on Figure 6.1. All nodes are constant rate servers, with unspecified arbitration between the flows.

6.3. STABILITY AND BOUNDS FOR A NETWORK WITH AGGREGATE SCHEDULING

183

Thus we are in the case where all nodes are strict service curve elements, with service curves of the form βm = λ C m . The method has two steps. First, we assume that there is a finite burstiness bound for all flows; using Section 6.2 we obtain some equations for computing these bounds. Second, we use the same equations to show that, under some conditions, finite bounds exist.

N o d e 0 F lo w 0

F lo w 3

N o d e 2

N o d e 1

Figure 6.1: A simple example with aggregate scheduling, used to illustrate the bounding method. There are three nodes numbered 0, 1, 2 and six flows, numbered 0, ..., 5. For i = 0, 1, 2, the path of flow i is i, (i+1) mod 3, (i+2) mod 3 and the path of flow i + 3 is i, (i + 2) mod 3, (i + 1) mod 3. The fresh arrival curve is the same for all flows, and is given by αi = γρ,σ . All nodes are constant rate, work conserving servers, with rate C. The utilization factor at all nodes is 6 Cρ .

F IRST STEP : INEQUATIONS FOR THE BOUNDS For any flow i and any node m ∈ i, define σim as the maximum backlog that this flow would generate in a constant rate server with rate ρi . By convention, the fresh inputs are considered as the outputs of a virtual node numbered −1. In this first step, we assume that σim is finite for all i and m ∈ i. By applying Corollary 6.2.2 we find that for all i and m ∈ i: ⎧ 0 ⎨ σi ≤ σ i ⎩ σ m = σ predi (m) + ρ i i i

jm,j =i

C−

predj (m)

σj

(6.6)

jm,j=i ρj

where predi (m) is the predecessor of node m. If m is the first node on the path of flow i, we set by convention predi (m) = −1 and σi−1 = σi . Now put all the σim , for all (i, m) such that m ∈ i, into a vector x with one column and n rows, for some appropriate n. We can re-write Equation (6.6) as x ≤ Ax + a

(6.7)

where A is an n × n, non-negative matrix and a is a non-negative vector depending only on the known quantities σi . The method now consists in assuming that the spectral radius of matrix A is less than 1. In that case the power series I + A + A2 + A3 + ... converges and is equal to (I − A)−1 , where I is the n × n identity matrix. Since A is non-negative, (I − A)−1 is also non-negative; we can thus multiply Equation (6.6) to the left by (I − A)−1 and obtain: x ≤ (I − A)−1a

(6.8)

which is the required result, since x describes the burstiness of all flows at all nodes. From there we can obtain bounds on delays and backlogs.


184

Let us apply this step to our network example. By symmetry, we have only two unknowns x and y, defined as the burstiness after one and two hops:

Equation (6.6) becomes

x = b00 = b11 = σ22 = b03 = b24 = b15 y = b10 = b21 = σ20 = b23 = b14 = b05

ρ (σ + 2x + 2y) x ≤ σ + C−5ρ ρ y ≤ x + C−5ρ (2σ + x + 2y)

ρ Define η = C−5ρ ; we assume that the utilization factor is less than 1, thus 0 ≤ η < 1. We can now write Equation (6.7) with

x 2η 2η σ(1 + η) x = , A= , a = y 1 + η 2η 2ση

Some remnant from linear algebra, or a symbolic computation software, tells us that −1

(I − A)

=

1−2η 1−6η+2η 2 1+η 1−6η+2η 2

2η 1−6η+2η 2 1−2η 1−6η+2η 2

√ If η < 12 (3 − 7) ≈ 0.177 then (I − A)−1 is positive. This is the condition for the spectral radius of A to be less than 1. The corresponding condition on the utilization factor ν = 6ρ C is √ 8− 7 ≈ 0.564 ν 0, consider the virtual system made of the original network, where all sources are stopped at time τ . For this network the total number of bits in finite, thus we can apply the conclusion of step 1, and the burstiness terms are bounded by Equation (6.8). Since the right-handside Equation (6.8) is independent of τ , letting τ tend to +∞ shows the following. P ROPOSITION 6.3.1. With the notation in this section, if the spectral radius of A is less than 1, then the burstiness terms bm i are bounded by the corresponding terms in Equation (6.8). Back to the example of Figure 6.1, we find that if the utilization factor ν is less than 0.564, then the burstiness terms x and y are bounded by 2 x ≤ 2σ 18−33ν+16ν 36−96ν+57ν 2 18−18ν+ν 2 y ≤ 2σ 36−96ν+57ν 2 The aggregate traffic at any of the three nodes is γ6ρ,b -smooth with b = 2(σ + x + y). Thus a bound on delay is (see also Figure 6.2): σ 108 − 198ν + 91ν 2 b =2 d= C C 36 − 96ν + 57ν 2

6.4. STABILITY RESULTS AND EXPLICIT BOUNDS

185

100 80 60 40 20

0.2

0.4

0.6

0.8

1

Figure 6.2: The bound d on delay at any node obtained by the method presented here for the network of Figure 6.1 σ (namely, dC ), plotted as a function of the utilization factor. The thick (thin line). The graph shows d normalized by C σ line is a delay bound obtained if every flow is re-shaped at every output.

T HE CRITICAL LOAD FACTOR FOR THIS EXAMPLE For the network in this example, where we impose the constraint that all ρi are equal, we find νcri ≥ ν0 ≈ 0.564, which is much less than 1. Does it mean that no finite bound exists for ν0 ≤ ν < 1 ? The answer to this question is not clear. First, the ν0 found with the method can be improved if we express more arrival constraints. Consider our particular example: we have not exploited the fact that the fraction of input traffic to node i that originates from another node has to be λC -smooth. If we do so, we will obtain better bounds. Second, if we know that nodes have additional properties, such as FIFO, then we may be able to find better bounds. However, even so, the value of νcri seems to be unknown.

T HE PRICE FOR AGGREGATE SCHEDULING Consider again the example on Figure 6.1, but assume now that every flow is reshaped at every output. This is not possible with differentiated services, since there is no per-flow information at nodes other than access nodes. However, we use this scenario as a benchmark that illustrates the price we pay for aggregate scheduling. With this assumption, every flow has the same arrival curve at every node. Thus we can compute a service curve β1 for flow 1 (and thus for any flow) at every node, using Theorem 6.2.1; we find that β1 is the rate5σ . Thus a delay bound for flow at any node, including latency function with rate (C − 5ρ) and latency C−5ρ 6C the re-shaper, is h(α1 , α1 ⊗ β1 ) = h(α1 , β1 ) = C−5ρ for ρ ≤ C6 . Figure 6.2 shows this delay bound, compared to the delay bound we found if no reshaper is used. As we already know, we see that with perflow information, we are able to guarantee a delay bound for any utilization factor ≤ 1. However, note also that for relatively small utilization factors, the bounds are very close.

6.4

S TABILITY R ESULTS AND E XPLICIT B OUNDS

In this section we give strong results for two specific case. The former is for a unidirectional ring of aggregate servers (of any type, not necessarily FIFO or strict service curve). We show that for all rings, νcri = 1. The latter is for any topology, but with restrictions on the network type: packets are of fixed size and all links have the same bit rate.


186

6.4.1

T HE R ING IS S TABLE

The result was initially obtained in [77] for the case of a ring of constant rate servers, with all servers having the same rate. We give here a more general, but simpler form. A SSUMPTION AND N OTATION We take the same assumptions as in the beginning of Section 6.3 and assume in addition that the network topology is a unidirectional ring. More precisely: • The network is a unidirectional ring of M nodes, labelled 1, ..., M . We use the notation m ⊕ k = (m + k − 1) mod M + 1 and m k = (m − k − 1) mod M + 1, so that the successor of node m on the ring is node m ⊕ 1 and its predecessor is node m 1. • The route of flow i is (0, i.first, i.first ⊕ 1, ..., i.first ⊕ (hi − 1)) where 0 is a virtual node representing the source of flow i, i.first is the first hop of flow i, and hi is the number of hops of flow i. At its last hop, flow i exits the network. We assume that a flow does not wrap, namely, hi ≤ M . If hi = M , then the flow goes around theall ring, exiting at the same node it has entered. • Let bm = em rm and let b = m bm reflect the total latency of the ring. • For any node m let σ (m) = im σi . (m) and σ = Let σmax = maxM m=1 σ i σi . Note that σmax ≤ σ ≤ M σmax . (m) • Define η = minm (rm − ρ ). (m) (m) + . µ reflects the sum of the peak rate • Let ρ0 = i.first=m ρi and µ = maxM m=0 Cm − rm + ρ of transit links and the rates of fresh sources, minus the rate guaranteed to the aggregate of microflows. We expect high values of µ to give higher bounds. T HEOREM 6.4.1. If η > 0 (i.e. if the utilization factor is < 1) the backlog at any node of the unidirectional ring is bounded by µ M (M σmax + b) + σ + b η

The proof relies on the concept of chain of busy periods, combined with the time stopping P ROOF : method in Section 6.3.2. For a node m and a flow i, define Rim (t) as the cumulative amount of data of flow i at the output of node m. For m = 0, this defines the input function. Also define & ' xm (t) = (6.10) Ri0 (t) − Rim (t) im

thus xm (t) is the total amount of data that is present in the network at time t and will go through node m at some time > t. We also define the backlog at node m by qm (t) =

Rim1 (t) +

im,i.first =m

i.first=m

Ri0 (t) −

Rim (t)

im

Now obviously, for all time t and node m: qm (t) ≤ xm (t) and xm (t) ≤

M n=1

qn (t)

(6.11)

(6.12)


187

(Step 1) Assume that a finite bound X exists. Consider a time t and a node m that achieves the bound: xm (t) = X. We fix m and apply Lemma 6.4.1 to all nodes n. Call sn the time called s in the lemma. Since xn (sn ) ≤ X, it follows from the first formula in the lemma that (t − sn )η ≤ M σmax + b

(6.13)

By combining this with the second formula in the lemma we obtain qn (t) ≤ µ Now we apply Equation (6.12) and note that X≤M

M σmax + b (n) + bn + σ 0 η

M

(n) n=1 σ0

= σ, from which we derive

µ (M σmax + b) + σ + b η

(6.14)

(Step 2) By applying the same reasoning as in Section 6.3.2, we find that Equation (6.14) is always true. The theorem follows from Equation (6.11). L EMMA 6.4.1. For any nodes m, n (possibly with m = n), and for any time t there is some s such that xm (t) ≤ xn (s) − (t − s)η + M σmax + b (n) qn (t) ≤ (t − s)µ + bn + σ0 (n)

with σ0

P ROOF :

=

i.first=n σi .

By definition of the service curve property at node m, there is some s1 such that Rim (t) ≥ Rim1 (s1 ) + Ri0 (s1 ) + rm (t − s1 ) − bm im

im,i.first =m

i.first=m

which we can rewrite as

Rim (t) ≥ −A +

im

Ri0 (s1 ) + rm (t − s1 ) − bm

im

with

A=

&

' Ri0 (s1 ) − Rim−1 (s1 )

im,i.first =m

Now the condition {i m, i.first = m} implies that flow i passes through node m−1, namely, {i (m − 1)}. Furthermore, each element in the summation that constitutes A is nonnegative. Thus & ' A≤ Ri0 (s1 ) − Rim−1 (s1 ) = xm1 (s1 ) i(m−1)

Thus

Rim (t) ≥ −xm1 (s1 ) +

im

Ri0 (s1 ) + rm (t − s1 ) − bm

im

Now combining this with the definition of xm (t) in Equation (6.10) gives: & ' Ri0 (t) − Ri0 (s1 ) − rm (t − s1 ) + bm xm (t) ≤ xm1 (s1 ) + im

(6.15)


188

From the arrival curve property applied to all micro-flows i in the summation, we derive: xm (t) ≤ xm1 (s1 ) − (rm − ρ(m) )(t − s1 ) + σ (m) + bm and since rm − ρ(m) ≥ η and σ (m) ≤ σmax by definition of η and σmax , we have xm (t) ≤ xm1 (s1 ) − (t − s1 )η + σmax + bm We apply the same reasoning to node m 1 and time s1 , and so on iteratively until we reach node n backwards from m. We thus build a sequence of times s0 = t, s1 , s2 , ..., sj , ..., sk such that xmj (sj ) ≤ xm(j+1) (sj+1 ) − (t − sj+1 )η + σmax + bmj

(6.16)

until we have m k = n. If n = m we reach the same node again by a complete backwards rotation and k = M . In all cases, we have k ≤ M . By summing Equation (6.16) for j = 0 to k − 1 we find the first part of the lemma. Now we prove the second part. s = sk is obtained by applying the service curve property to node n and time sk−1 . Apply the service curve property to node n and time t. Since t ≥ sk−1 , we know from Proposition 1.3.2 on Page 19 that we can find some s ≥ s such that Rin (t) ≥ Rin−1 (s ) + Ri0 (s ) + rn (t − s ) − bn in

in,i.first =n

Thus

qn (t) ≤

i.first=n

' & n1 Ri (t) − Rin1 (s ) +

in,i.first =n

(Ri0 (t) − Ri0 (s )) − rn (t − s ) + bn

i.first=n

≤ (Cn − rn + ρ0 )(t − s ) + bn + σ0 (n)

(n)

≤ (t − s )µ + bn + σ0

(n)

the second part of the formula follows from s ≤ s . R EMARK :

A simpler, but weaker bound, is µ M (M σ + b) + σ + b η

or M

µ (M σmax + b) + M σmax + b η

(6.17)

T HE SPECIAL CASE IN [77]: Under the assumption that all nodes are constant rate servers of rate equal to 1 (thus Cm = rm = 1 and bm is the latency of the link m), the following bound is found in [77]: B1 =

M b + M 2 σmax +b η

(6.18)

In that case, we have µ ≤ 1 − η. By applying Equation (6.17), we obtain the bound M µb + M 2 µ + M η σmax +b B2 = η since µ≤1−η

(6.19)

and 0 < η ≤ 1, M ≤ we have B2 < B1 , namely, our bound is better than that in [77]. If there is equality in Equation (6.19) (namely, if there is a node that receives no transit traffic), then both bounds are equivalent when η → 0. M 2,


6.4.2

189

E XPLICIT B OUNDS FOR A H OMOGENEOUS ATM N ETWORK WITH S TRONG S OURCE R ATE C ONDITIONS

When analyzing a global network, we can use the bounds in Section 6.2.2, using the same method as in Section 2.4. However, as illustrated in [41], the bounds so obtained are not optimal: indeed, even for a FIFO ring, the method does not find a finite bound for all utilization factors less than 1 (although we know from Section 6.4.1 that such finite bounds exist). In this section we show in Theorem 6.4.2 some partial result that goes beyond the per-node bounds in Section 6.2.2. The result was originally found in [15, 52, 83]. Consider an ATM network with the assumptions as in Section 6.3, with the following differences • Every link has one origin node and one end node. We say that a link f is incident to link e if the origin node of link e is the destination node of link f . In general, a link has several incident links. • All packets have the same size (called cell). All arrivals and departures occur at integer times (synchronized model). All links have the same bit rate, equal to 1 cell per time unit. The service time for one cell is 1 time unit. The propagation times are constant per link and integer. • All links are FIFO. P ROPOSITION 6.4.1. For a network with the above assumption, the delay for a cell c arriving at node e over incident link i is bounded by the number of cells arriving on incident links j = i during the busy period, and that will depart before c. P ROOF : Call R (t) (resp. Rj (t), R(t))the output flow (resp. input arriving on link j, total input flow). Call d the delay for a tagged cell arriving at time t on link i. Call Aj the number of cells arriving on link j up to time t that will depart before the tagged cell, and let A = j Aj . We have d = A − R (t) ≤ A − R(s) − (t − s) where s is the last time instant before the busy period at t. We can rewrite the previous equation as [Aj − Rj (s)] + [Ai (t) − Ri (s)] − (t − s) d≤ j =i

Now the link rates are all equal to 1, thus Ai − Ri (s) ≤ t − s and d≤ [Aj − Rj (s)] j =i

An “Interference Unit” is defined as a set (e, {j, k}) where e is a link, {j, k} is a set of two distinct flows that each have e on their paths, and that arrive at e over two different incident links (Figure 6.3). The Route Interference Number (RIN) of flow j is the number of interference units that contain j. It is thus the number of other flows that share a common sub-path, counted with multiplicity if some flows share several distinct sub-paths along the same path. The RIN is used to define a sufficient condition, under which we prove a strong bound. D EFINITION 6.4.1 (S OURCE R ATE C ONDITION ). The fresh arrival curve constraint (at network boundary) for flow j is the stair function vR+1,R+1 , where R is the RIN of flow j. The source rate condition is equivalent to saying that a flow generates at most one cell in any time interval of duration RIN + 1.


190 f lo w i2

f lo w j n o d e l

n o d e h

n o d e g

n o d e f

n o d e e

n o d e k f lo w i1

Figure 6.3: The network model and definition of an interference unit. Flows j and i2 have an interference unit at node f . Flows j and i1 have an interference unit at node l and one at node g. T HEOREM 6.4.2. If the source rate condition holds at all sources, then 1. The backlog at any node is bounded by N − maxi Ni , where Ni is the number of flows entering the node via input link i, and N = i Ni . 2. The end-to-end queuing delay for a given flow is bounded by its RIN. 3. There is at most one cell per flow present during any busy period. The proof of item 3 involves a complex analysis of chained busy periods, as does the proof of Theorem 6.4.1. It is given in a separate section. Item 3 gives an intuitive explanation of what happens: the source rate condition forces sources to leave enough spacing between cells, so that two cells of the same flow do not interfere, in some sense. The precise meaning of this is given in the proof. Items 1 and 2 derive from item 3 by a classical network calculus method (Figure 6.6). P ROOF OF T HEOREM 6.4.2 As a simplification, we call “path of a cell“ the path of the flow of the cell. Similarly, we use the phrase “interference unit of c” with the meaning of interference unit of the flow of c. We define a busy period as a time interval during which the backlog for the flow at the node is always positive. We now introduce a definition (super-chain) that will be central in the proof. First we use the following relation: D EFINITION 6.4.2 (“D ELAY C HAIN ” [15]). For two cells c and d, and for some link e, we say that c e d if c and d are in the same busy period at e and c leaves e before d. Figure 6.4 illustrates the definition. D EFINITION 6.4.3 (S UPER -C HAIN [15]). Consider a sequence of cells c = (c0 , ..., ci , ..., ck ) and a sequence of nodes f = (f1 , ..., fk ). We say that (c, f ) is a super-chain if • f1 , ..., fk are all on the path P of cell c0 (but not necessarily consecutive) • ci−1 fi ci for i = 1 to k. • the path of cell ci from fi to fi+1 is a sub-path of P We say that the sub-path of c0 that spans from node f1 to node fk is the path of the super-chain. D EFINITION 6.4.4 (S EGMENT I NTERFERING WITH A S UPER -C HAIN ). For a given super-chain, we call “segment” a couple (d, P ) where P is a sub-path of the path of the super-chain, d is a cell whose path also has P as a sub-path, and P is maximal (namely, we cannot extend P to be a common sub-path of both d and the super-chain). We say that the segment (d, P ) is interfering with super-chain (c, f ) if there is some i on P such that d fi ci .


191

c e ll d c e ll c 1

c e ll c

l h

g

c e ll c

c e ll d

2

e 2

f c e ll c 1

Figure 6.4: A time-space diagram illustrating the definitions of d g c1 and c1 f c2 . Time flows downwards. Rectangles illustrate busy periods. L EMMA 6.4.2. Let (c, f ) be a super-chain. Let s0 be the arrival time of cell c0 at link f1 and sk the departure time of cell ck from link fk . Then sk − s0 ≤ R1,k + T1,k , where R1,k is the total number of segments interfering with (c, f ) and T1,k is the total transmission and propagation time on the path of the super-chain. P ROOF : Consider first some node fj on the super-chain. Let sj−1 (resp. tj ) be the arrival time of cell cj−1 (resp. cj ) at the node. Let tj−1 (resp. sj ) be the departure time of cell cj−1 (resp. cj ) (Figure 6.5). Let vj be the last time slot before the busy period that tj is in. By hypothesis, vj + 1 ≤ sj−1 . Also define Bj

v B

j

j

C e ll c

s

j- 1

A

0

j- 1

j

t

t j - 1 j

C e ll c j

s j t im e

Figure 6.5: The notation used in the proof of Lemma 6.4.2. (resp. Bj0 ) as the set of segments (d, P ) where d is a cell arriving at the node after time vj on a link incident


192

to the path of the super-chain (resp. on the path of the super-chain) and that will depart no later than cell cj , and where P is the maximal common sub-path for d and the super-chain that fj is in. Also define A0j as the subset of those segments in Bj0 for which the cell departs after cj−1 . Let Bj (resp. Bj0 , A0j ) be the number of elements in Bj (resp. Bj0 , A0j ), see Figure 6.5. Since the rate of all incident links is 1, we have Bj0 − A0j ≤ sj−1 − vj Also, since the rate of the node is 1, we have: sj − vj = Bj + Bj0 Combining the two, we derive sj − sj−1 = Bj + Bj0 − (sj−1 − vj ) ≤ Bj + A0j

(6.20)

By iterative application of Equation (6.20) from j = 1 to k, we obtain sk − s0 ≤

k

(Bj + A0j ) + T1,k

j=1

Now we show that all sets in the collection {Bj , A0j , j = 1 to k} are two-by-two disjoint. Firstly, if (d, P ) ∈ Bj then fj is the first node of P thus (d, P ) cannot be in some other Bj with j = j . Thus the Bj are two-by-two disjoint. Second, assume (d, P ) ∈ Bj and (d, P ) ∈ A0j . It is obvious from their definitions that, for a fixed j, Bj and A0j are disjoint; thus j = j . Since fj is the first node of P and j is on P , it follows that j < j . Now d leaves fj before cj and leaves fj after cj −1 , which contradicts the FIFO assumption. Thus(the Bj and A0j are two-by-two disjoint. The same reasoning shows that it is not possible that (d, P ) ∈ Aj Aj with j < j . Now, by definition, every segment in either Bj or A0j is an interfering segment. Thus k (Bj + A0j ) ≤ R1,k j=1

. P ROPOSITION 6.4.2. Assume the source rate condition holds. Let (c, f ) be a super-chain. 1. For every interference unit of c0 there is at most one cell interfering with the super-chain. 2. ck does not belong to the same flow as c0 . P ROOF : Define the time of a super-chain as the exit time for the last cell ck on the last node fk . We use a recursion on the time t of the super-chain. If t = 1, the proposition is true because any flow has at most one cell on a link in one time slot. Assume now that the proposition holds for any super-chain with time ≤ t − 1 and consider a super-chain with time t. First, we associate an interference unit to any segment (d, P ) interfering with the sub-chain, as follows. The paths of d and c0 may share several non contiguous sub-paths, and P is one of them. Call f the first node of P . To d we associate the interference unit (f, {j0 , j}), where j0 (resp. j) is the flow of c0 (resp. d). We now show that this mapping is injective. Assume that another segment (d , P ) = (d, P ) is associated with the same interference unit (f, {j0 , j}). Without loss of generality, we can assume that d was emitted


193

before d . d and d belong to the same flow j, thus, since P and P are maximal, we must have P = P . By hypothesis, have an interference with the super-chain at a node on P . Let fl be a node on the super-chain and on P such that d fl cl . If d leaves node fl before cl , then d fl d , and thus ((d, d ), (fl )) is a super-chain. Since d is an interfering cell, necessarily, it must leave node fl before t, thus the proposition is true for super-chain ((d, d ), (fl )), which contradicts item 2. Thus d must leave node fl after cell cl . But there is some other index m ≤ k such that d fm cm , thus cell d leaves node fm before cell cm . Define l as the smallest index with l < l ≤ m such that d leaves node fl after cell cl −1 and before cl . Then ((d, cl , ..., cl −1 , d ), (fl , .., fl )) is a super-chain with time ≤ t − 1 which would again contradict item 2 in the proposition. Thus, in all cases we have a contradiction, the mapping is injective, and item 1 is shown for the super-chain. Second, let us count a bound on the maximum queuing delay of cell c0 . Call u0 its emission time, P0 the subpath of c0 from its source up to, but excluding, node f1 , and T the total transmission and propagation time for the flow of c0 . The transmission and propagation time along P0 is thus T − T1,k . By Proposition 6.4.1, the queuing delay of c0 at a node f on P0 is bounded by the number of cells d f c0 that arrive on a link not on P0 . By the same reasoning as in the previous paragraph, there is at most one such cell d per interference unit of c0 at f . Define R as the number of interference units of the flow of c0 on P1 . We have thus s0 ≤ u0 + R + T − T1,k

(6.21)

Similarly, from Lemma 6.4.2, we have sk ≤ s0 + R1,k + T1,k Call R the number of interference units of the flow of c0 on the path of the super-chain. It follows from the first part of the proof that R1,k ≤ R , thus sk ≤ s0 + R + T1,k Combining with Equation (6.21) gives sk ≤ u0 + R + R + T

(6.22)

Now by the source condition, if ck belongs to the flow of c0 , its emission time u must satisfy u ≥ u0 + R + R + 1 and thus sk ≥ u0 + R + R + 1 + T which contradicts Equation (6.22). This shows that the second item of the proposition must hold for the super-chain. P ROOF OF T HEOREM 6.4.2: Item 3 follows from Proposition 6.4.2, since if there would be two cells d, d of the same flow in the same busy period, then ((d, d ), (e)) would be a super-chain. Now we show how items 1 and 2 derive from item 3. Call αi∗ (t) the maximum number of cells that may ever arrive on incident link i during t time units inside a busy period. Since λ1 is a service curve for node e, the backlog B at node e is bounded by ) B ≤ sup t≥0

I i=1

* αi∗ (t) − t


194 Now by item 3, αi∗ (t) ≤ Ni and thus

αi∗ (t) ≤ αi (t) := min[Ni , t] Thus

) B ≤ sup t≥0

I

* αi (t) − t

i=1

Now define a renumbering of the Ni ’s such that N(1) ≤ N(2) ≤ ... ≤ N(I) . The function i αi (t) − t is continuous and has a derivative at all points except the N(i) ’s (Figure 6.6). The derivative changes its sign at N(I) (=max1≤i≤I (Ni )) thus the maximum is at N(I) and its value is N − N(I) , which shows item 1.

a (t ) N N - N

N

(1 )

N

N

(2 )

t

(I )

(I )

Figure 6.6: Derivation of a backlog bound. From Item 1, the delay at a node is bounded by the number of interference units of the flow at this node. This shows item 2.

6.5


In [52], a stronger property is shown than Theorem 6.4.2: Consider a given link e and a subset A of m connections that use that link. Let n be a lower bound on the number of route interferences that any connection in the subset will encounter after this link. Then over any time interval of duration m + n, the number of cells belonging to A that leave link e is bounded by m. It follows from item 1 in Theorem 6.4.2 that a better queuing delay bound for flow j is: δ(j) =

e such that e∈j

min

i such that 1≤i≤I(e)

(N (e) − Ni (e))

where I(e) is the number of incident links at node e, Ni (e) is the number of flows entering node e on link i, and N = i = 1I(e) Ni (e). In other words, the end-to-end queuing delay is bounded by the sum of the minimum numbers of route interference units for all flows at all nodes along the path of a flow. For asymmetric cases, this is less than the RIN of the flow.

6.6. EXERCISES

6.6

195

E XERCISES

E XERCISE 6.1. Consider the same assumptions as in Section 6.4.1 but with a linear network instead of a ring. Thus node m feeds node m + 1 for m = 1, ..., M − 1; node 1 receives only fresh traffic, whereas all traffic exiting node M leaves the network. Assume that all service curves are strict. Find a bound which is finite for ν ≤ 1. Compare to Theorem 6.4.1. E XERCISE 6.2. Consider the same assumptions as in Theorem 6.4.2. Show that the busy period duration is bounded by N . E XERCISE 6.3. Consider the example of Figure 6.1. Apply the method of Section 6.3.2 but express now that the fraction of input traffic to node i that originates from another node must have λC as an arrival curve . What is the upper-bound on utilization factors for which a bound is obtained ? E XERCISE 6.4. Can you conclude anything on νcri from Proposition 2.4.1 on Page 90 ?

196


C HAPTER 7

A DAPTIVE AND PACKET S CALE R ATE G UARANTEES

7.1

I NTRODUCTION

In Chapter 1 we defined a number of service curve concepts: minimum service curve, maximum service curve and strict service curves. In this chapter we go beyond and define some concepts that more closely capture the properties of generalized processor sharing (GPS). We start by a motivating section, in which we analyze some features of service curves or Guaranteed Rate node that do not match GPS. Then we provide the theoretical framework of packet scale rate guarantee (PSRG); it is a more complex node abstraction than Guaranteed Rate, which better captures some of the properties of GPS. A major difference is the possibility to derive information on delay when the buffer size is known – a property that is not possible with service curve or guaranteed rate. This is important for low delay services in the internet. PSRG is used in the definition of the Internet Expedited Forwarding service. Just like GR is the max-plus correspondant of the min-plus concept of service curve, PSRG is the max-plus correspondant of adaptive service curves. These were first proposed in Okino’s dissertation in [62] and by Agrawal, Cruz, Okino and Rajan in [1]. We explain the relationship between the two and give practical applications to the concatenation of PSRG nodes. In the context of differentiated services, a flow is an aggregate of a number of micro-flows that belong to the same service class. Such an aggregate may join a router by different ports, and may follow different paths inside the router. It follows that it can generally not be assumed that a router is FIFO per flow. This is why the definition of PSRG (like GR) does not assume the FIFO property. In all of this chapter, we assume that flow functions are left-continuous, unless stated otherwise.

7.2

L IMITATIONS OF THE S ERVICE C URVE AND GR N ODE A BSTRAC TIONS

The definition of service curve introduced in Section 1.3 is an abstraction of nodes such as GPS and its practical implementations, as well as guaranteed delay nodes. This abstraction is used in many situations, described all along this book. However, it is not always sufficient. 197

CHAPTER 7. ADAPTIVE AND PACKET SCALE RATE GUARANTEES

198

Firstly, it does not provide a guarantee over any interval. Consider for example a node offering to a flow R(t) the service curve λC . Assume R(t) = B for t > 0, so the flow has a very large burst at time 0 and then stops. A possible output is illustrated on Figure 7.1. It is perfectly possible that there is no output during the time interval (0, B− C ], even though there is a large backlog. This is because the server gave a higher service than the minimum required during some interval of time, and the service property allows it to be lazy after that. B B - e

R (t ) R * (t )

C t

Figure 7.1: The service curve property is not sufficient. Secondly, there are case where we would like to deduce a bound on the delay that a packet will suffer given the backlog that we can measure in the node. This is used for obtaining bounds in FIFO systems with aggregate scheduling. In Chapter 6 we use such a property for a constant delay server with rate C: given that the backlog at time t is Q, the last bit present at time t will depart before within a time of Q C . If we assume instead that the server has a service curve λC , then we cannot draw such a conclusion. Consider for example Figure 7.1: at time t > 0, the backlog, , can be made arbitrily small, whereas the delay B− C −t can be made arbitrarily large. The same limitation applies to the concept of Guaranteed Rate node. Indeed, the example in Figure 7.1 could very well be for GR node. The main issue here is that a GR node, like a service curve element, may serve packets earlier than required. A possible fix is the use of strict service curve, as defined in Definition 1.3.2 on Page 21. Indeed, it follows from the next section (and can easily be shown independently) that if a FIFO node offers a strict service curve β, then the delay at time t is bounded by β −1 (Q(t)), where Q(t) is the backlog at time t, and β −1 is the pseudo-inverse (Definition 3.1.7 on Page 108). We know that the GPS node offers to a flow a strict service curve equal of the form λR . However, we cannot model delay nodes with a strict service curve. Consider for example a node with input R(t) = t, which delays all bits by a constant time d. Any interval [s, t] with s ≥ d is within a busy period, thus if the node offers a strict service curve β to the flow, we should have β(t − s)(t − s), and can be arbitrarily small. Thus, the strict service curve does not make much sense for a constant delay node.

7.3 7.3.1

PACKET S CALE R ATE G UARANTEE D EFINITION OF PACKET S CALE R ATE G UARANTEE

In Section 2.1.3 on Page 70 we have introduced the definition of guaranteed rate scheduler, which is the practical application of rate latency service curves. Consider a node where packets arrive at times a1 ≥ 0, a2 , ... and leave at times d1 , d2 , .... A guaranteed rate scheduler, with rate r and latency e requires that

7.3. PACKET SCALE RATE GUARANTEE

199

di ≤ fi + e, where fi is defined iteratively by f0 = 0 and fi = max{ai , fi−1 }+

li r

where li is the length of the ith packet. A packet scale rate guarantee is similar, but avoids the limitations of the service curve concept discussed in Section 7.2. To that end, we would like that the deadline fi is reduced whenever a packet happens to be served early. This is done by replacing fi−1 in the previous equation by min{fi , di }. This gives the following definition. D EFINITION 7.3.1 (PACKET S CALE R ATE G UARANTEE ). Consider a node that serves a flow of packets numbered i = 1, 2, .... Call ai , di , li the arrival time, departure time, and length in bits for the ith packet, in order of arrival. Assume a1 ≥ 0.We say that the node offers to the flow a packet scale rate guarantee with rate r and latency e if the departure times satisfy di ≤ fi + e where fi is defined by:

f0 = d0 = 0 fi = max {ai , min (di−1 , fi−1 )} +

li r

for all i ≥ 1

See Figure 7.2 and Figure 7.3 for an illustration of the definition.

IQ IQ PD[^DQ PLQ>GQ IQ @`/Q U PD[^DQ PLQ>GQ IQ @`/Q U /Q U IQ GQ DQ

IQ

GQ

/Q U DQ

IQ GQ IQ /Q U

DQ GQ IQ IQ GQ

Figure 7.2: Definition of PSRG. T HEOREM 7.3.1. A PSRG node with rate r and latency e is GR(r, e).

GQ

(7.1)


200

3DFNHW6FDOH5DWH*XDUDQWHH

IQ IQ PD[^DQ PLQ>GQ IQ @`/Q U PD[^DQ PLQ>GQ IQ @`/Q U /Q U DQ GQ IQ IQ GQ

*XDUDQWHHG5DWH

IQ IQ PD[^DQ IQ @`/Q U PD[^DQ IQ @`/Q U /Q U DQ GQ IQ

IQ GQ

Figure 7.3: Difference between PSRG and GR when packet n − 1 leaves before fn . P ROOF :

Follows immediately from the definition.

Comment. It follows that a PSRG node enjoys all the properties of a GR node. In particular: • Delay bounds for input traffic with arrival curves can be obtained from Theorem 2.1.4. • PSRG nodes have a rate latency service curve property (Theorem 2.1.3) that can be used for buffer dimensioning. We now obtain a characterization of packet scale rate guarantee that does not contain the virtual finish times fn . It is the basis for many results in this chapter. We start with an expansion of the recursive definition of packet scale rate guarantee, L EMMA 7.3.1 (M IN - MAX EXPANSION OF PSRG). Consider three arbitrary sequences of non-negative numbers (an )n≥1 , (dn )n≥0 , and (mn )n≥1 , with d0 = 0. Define the sequence (fn )n≥0 , by f0 = 0 fn = max [an , min (dn−1 , fn−1 )] + mn for n ≥ 1 Also define

Anj = aj + mj + ... + mn for 1 ≤ j ≤ n Djn = dj + mj+1 + ... + mn for 0 ≤ j ≤ n − 1

For all n ≥ 1, we have fn = min [ max(Ann , Ann−1 , ..., An1 ), max(Ann , Ann−1 ..., An2 , D1n ), ... max(Ann , Ann−1 ..., Anj+1 , Djn ),

7.3. PACKET SCALE RATE GUARANTEE

201 ... n max(Ann , Ann−1 , Dn−2 ), n max(Ann , Dn−1 )

] The proof is long and is given in a separate section (Section 7.7); it is based on min-max algebra. Comment: The expansion in Lemma 7.3.1 can be interpreted as follows. The first term max(Ann , Ann−1 , ..., An1 ) corresponds to the guaranteed rate clock recursion (see Theorem 2.1.2). The following terms have the effect of reducing fn , depending on the values of dj . We now apply the previous lemma to packet scale rate guarantee and obtain the required characterization without the virtual finish times fn : T HEOREM 7.3.2. Consider a system where packets are numbered 1, 2, ... in order of arrival. Call an , dn the arrival and departure times for packet n, and ln the size of packet n. Define by convention d0 = 0. The packet scale rate guarantee with rate r and latency e is equivalent to: For all n and all 0 ≤ j ≤ n − 1, one of the following holds lj+1 + ... + ln (7.2) dn ≤ e + dj + r or there is some k ∈ {j + 1, ..., n} such that dn ≤ e + ak +

lk + ... + ln r

(7.3)

The proof is also given in Section 7.7. It is a straightforward application of Lemma 7.3.1. Comment 1: The original definition of EF in [42] was based on the informal intuition that a node guarantees to the EF aggregate a rate equal to r, at all time scales (this informal definition was replaced by PSRG). Theorem 7.3.2 makes the link to the original intuition: a rate guarantee at all time scales means that either Equation (7.2) or Equation (7.3) must hold. For a simple scheduler, the former means that dj , dn are in the same backlogged period; the latter is for the opposite case, and here ak is the beginning of the backlogged period. But note that we do not assume that the PSRG node is a simple scheduler; as mentioned earlier, it may be any complex, non work conserving node. It is a merit of the abstract PSRG definition to avoid using the concept of backlogged period, which is not meaningful for a composite node [13, 5]. Comment 2: In Theorem 2.1.2 we give a similar result for GR nodes. It is instructive to compare both in the case of a simple scheduler, where the interpretation in terms of backlogged period can be made. Let us assume the latency term is 0, to make the comparison simple. For such a simple scheduler, PSRG means that during any backlogged period, the scheduler guarantees a rate at least equal to r. In contrast, and again for such simple schedulers, GR means that during the backlogged period starting at the first packet arrival that finds the system empty (this is called “busy period” in queuing theory), the average rate of service is at least r. GR allows the scheduler to serve some packets more quickly than at rate r, and take advantage of this to serve other packets at a rate smaller than r, as long as the overall average rate is at least r. PSRG does not allow such a behaviour. A special case of interest is when e = 0. D EFINITION 7.3.2. We call minimum rate server, with rate r, a PSRG node for with latency e = 0. For a minimum rate server we have d0 = 0 di ≤ max {ai , di−1 } +

li r

for all i ≥ 1

(7.4)

Thus, roughly speaking, a minimum rate server guarantees that during any busy period, the instantaneous output rate is at least r. A GPS node with total rate C and weight wi for flow i is a minimum rate server for flow i, with rate ri = φi Cφj . j


202

7.3.2

P RACTICAL R EALIZATION OF PACKET S CALE R ATE G UARANTEE

We show in this section that a wide variety of schedulers provide the packet scale rate guarantee. More schedulers can be obtained by using the concatenation theorem in the previous section. A simple but important realization is the priority scheduler. P ROPOSITION 7.3.1. Consider a non-preemptive priority scheduler in which all packets share a single FIFO queue with total output rate C. The high priority flow receives a packet scale rate guarantee with rate C and latency e = lmax C , where lmax is the maximum packet size of all low priority packets. P ROOF :

By Proposition 1.3.7, the high priority traffic receives a strict service curve βr,c .

We have already introduced in Section 2.1.3 schedulers that can be thought of as derived from GPS and we have modeled their behaviour with a rate-latency service curve. In order to give a PSRG for such schedulers, we need to define more. D EFINITION 7.3.3 (PSRG ACCURACY OF A SCHEDULER WITH RESPECT TO RATE r). Consider a scheduler S and call di the time of the i-th departure. We say that the PSRG accuracy of S with respect to rate r is (e1 , e2 ) if there is a minimum rate server with rate r and departure times gi such that for all i gi − e1 ≤ di ≤ gi + e2

(7.5)

We interpret this definition as a comparison to a hypothetical GPS reference scheduler that would serve the same flows. The term e2 determines the maximum per-hop delay bound, whereas e1 has an effect on the jitter at the output of the scheduler. For example, it is shown in [6] that WF2 Q satisfies e1 (WF2 Q) = lmax /r, e2 (WF2 Q) = lmax /C, where lmax is maximum packet size and C is the total output rate. In contrast, for PGPS [64] e2 (PGPS) = e2 (WF2 Q), while e1 (PGPS) is linear in the number of queues in the scheduler. This illustrates that, while WF2 Q and PGPS have the same delay bounds, PGPS may result in substantially burstier departure patterns. T HEOREM 7.3.3. If a scheduler satisfies Equation (7.5), then it offers the packet scale rate guarantee with rate r and latency e = e1 + e2 . The proof is in Section 7.7.

7.3.3

D ELAY F ROM BACKLOG

A main feature of the packet scale rate guarantee definition is that it allows to bound delay from backlog. For a FIFO node, it could be derived from Theorem 7.4.3 and Theorem 7.4.5. But the important fact is that the bound is the same, with or without FIFO assumption. T HEOREM 7.3.4. Consider a node offering the Packet Scale Rate Guarantee with rate r and latency e, not necessarily FIFO. Call Q the backlog at time t. All packets that are in the system at time t will leave the system no later than at time t + Q/r + e, The proof is in Section 7.7. Application to Differentiated Services Consider a network of nodes offering the EF service, as in Section 2.4.1. Assume node m is a PSRG node with rate rm and latency em . Assume the buffer size at node m is limited to Bm . A bound D on delay at node m follows directly D=

Bm + em rm

7.4. ADAPTIVE GUARANTEE

203

1

GHOD\

0.8

0%

0.6 K H 078U U 0EV & 5

0.4 0.2

0.05

0.1

0.15

0.2

0.25

0% 0% α

Figure 7.4: End to end delay bound versus the utilization factor α for an infinite buffer (left curve) and buffers sizes of , σi = 100B and ρi = 32kb/s for 1MB (top), 0.38MB (middle) and 0.1MB (bottom). There are h = 10 hops, em = 2 1500B rm all flows, rm = 149.760Mb/s. Compare to the bound in Theorem 2.4.1: this bound is valid for all utilization levels and is independent of traffic load. Figure 7.4 shows a numerical example. However, forcing a small buffer size may cause some packet loss. The loss probability can be computed if we assume in addition that the traffic at network edge is made of stationary, independent flows [58].

7.4 7.4.1

A DAPTIVE G UARANTEE D EFINITION OF A DAPTIVE G UARANTEE

Much in the spirit of PSRG, we know introduce a stronger service curve concept, called adaptive guarantee, that better captures the properties of GPS [62, 1], and helps finding concatenation properties for PSRG. Before giving the formula, we motivate it on three examples. Example 1. Consider a node offering a strict service curve β. Consider some fixed, but arbitrary times s < t. Assume that β is continuous. If [s, t] is within a busy period, we must have R∗ (t) ≥ R∗ (s) + β(t − s) Else, call u the beginning of the busy period at t. We have R∗ (t) ≥ R(u) + β(t − u) thus in all cases

R∗ (t) ≥ (R∗ (s) + β(t − s)) ∧ inf (R(u) + β(t − u)) u∈[s,t]

Example 2. Consider a node that guarantees a virtual delay ≤ d. If t − s ≤ d then trivially R∗ (t) ≥ R∗ (s) + δd (t − s) and if t − s > d then the virtual delay property means that R∗ (t) ≥ R(t − d) = inf (R(u) + δd (t − u)) u∈[s,t]

(7.6)


204

thus we have the same relation as in Equation (7.6) with β = δd . Example 3. Consider a greedy shaper with shaping function σ (assumed to be a good function). Then R∗ (t) = inf [R(u) + σ(t − u)] u≤t

Breaking the inf into u < s and u ≥ s gives R∗ (t) = inf [R(u) + σ(t − u)] ∧ inf [R(u) + σ(t − u)] u 1. Assume system i offers the packet scale rate guarantee with rate ri and latency system offers the packet scale rate guarantee with rate r = mini=1,...,n ri and latency ri . The global e = i=1,...,n ei + i=1,...,n−1 Lmax ri . P ROOF : By Theorem 7.4.5–(2), we can decompose system i into a concatenation Si , Pi , where Si offers the adaptive guarantee βri ,ei and Pi is a packetizer. Call S the concatenation

S1 , P1 , S2 , P2 , ..., Sn−1 , Pn−1 , Sn

By Theorem 7.4.5–(2), S is FIFO. By Theorem 7.4.4, it provides the adaptive guarantee βr,e . By Theorem 7.4.5–(1), it also provides the packet scale rate guarantee with rate r and latency e. Now Pn does not affect the finish time of the last bit of every packet. A Composite Node We analyze in detail one specific example, which often arises in practice when modelling a router. We consider a composite node, made of two components. The former (“variable delay component”) imposes to packets a delay in the range [δmax − δ, δmax ]. The latter is FIFO and offers to its input the packet scale rate guarantee, with rate r and latency e. We show that, if the variable delay component is known to be FIFO, then we have a simple result. We first give the following lemma, which has some interest of its own. L EMMA 7.5.1 (VARIABLE D ELAY AS PSRG). Consider a node which is known to guarantee a delay ≤ δmax . The node need not be FIFO. Call lmin the minimum packet size. For any r > 0, the node offers the + packet scale rate guarantee with latency e = [δmax − lmin r ] and rate r. Proof. With the standard notation in this section, the hypothesis implies that dn ≤ an + δmax for all n ≥ 1. Define fn by , thus dn − fn ≤ δmax − lmin ≤ [δmax − lmin ]+ . Equation (7.1). We have fn ≥ an + lrn ≥ an + lmin r r r

We will now apply known results on the concatenation of FIFO elements and solve the case where the variable delay component is FIFO.

7.5. CONCATENATION OF PSRG NODES

207

T HEOREM 7.5.2. (Composite Node with FIFO Variable Delay Component) Consider the concatenation of two nodes. The former imposes to packets a delay ≤ δmax . The latter offers the packet scale rate guarantee to its input, with rate r and latency e. Both nodes are FIFO. The concatenation of the two nodes, in any order, offers the packet scale rate guarantee with rate r and latency e = e + δmax . Proof. Combine Theorem 7.4.2 with Lemma 7.5.1: for any r ≥ r, the combined node offers the packet scale guarantee with rate min . Define fn for all n by Equation (7.1). Consider some fixed but arbitrary n. We have r and latency e(r ) = e + δmax + lmaxr−l dn − fn ≤ e(r ), and this is true for any r ≥ r. Let r → +∞ and obtain dn − fn ≤ inf r ≥r e(r ) = e + δmax as required.

7.5.2

C ONCATENATION OF NON FIFO PSRG N ODES

In general, we cannot say much about the concatenation of non FIFO PSRG nodes. We analyze in detail composite node described above, but now the delay element is non FIFO. This is a frequent case in practice. The results are of interest for modelling a router. The also serve the purpose of showing that the results in Theorem 7.5.1 do not hold here. To obtain a result, we need to an arrival curve for the incoming traffic. This is because some packets may take over some other packets in the non-FIFO delay element (Figure 7.5); an arrival curve puts a bound on this. n o n F I F O , d e la y d

a k

l1

1

a 1

a 2

a 3

l2 l3

b

F I F O , P S R G k

2 b 2 = b

P 2 P 1 b

1

= b

(2 )

b 4

= b

(3 )

b 3

= b

(4 )

P 3

l4 a 4

(1 )

P 4

d k

P 2

d 2 = d

(1 )

d 1

= d

(2 )

d 4

= d

(3 )

3

= d

(4 )

P 1 P 4 P 3 d

Figure 7.5: Composite Node with non-FIFO Variable Delay Component. Packet n arrives at times an at the first component, at time bn at the second component, and leaves the system at time dn . Since the first component is not FIFO, overtaking may occur; (k) is the packet number of the kth packet arriving at the second component.

T HEOREM 7.5.3. (Composite Node with non-FIFO Variable Delay Component) Consider the concatenation of two nodes. The first imposes to packets a delay in the range [δmax −δ, δmax ]. The second is FIFO and offers the packet scale rate guarantee to its input, with rate r and latency e. The first node is not assumed to be FIFO, so the order of packet arrivals at the second node is not the order of packet arrivals at the first one. Assume that the fresh input is constrained by a continuous arrival curve α(·). The concatenation of the


208

two nodes, in this order, satisfies the packet scale rate guarantee with rate r and latency e = e + δmax + min min{supt≥0 [ α(t+δ)−l − t], r α(t)+α(δ)−2lmin sup0≤t≤δ [ − t]} r

(7.10)

The proof is long, and is given in Section 7.7. Figures 7.6 to 7.8 show numerical applications when the arrival curve includes both peak rate and mean rate constraints. Special Case : For α(t) = ρt + σ, a direct computation of the suprema in Theorem 7.5.3 gives: min e = e + δmax + ρδ+σ−l r ρδ+σ−l min e = e + δmax − δ + 2 r

if ρ ≤ r then else

The latency of the composite node has a discontinuity equal to σ/r at ρ = r. It may seem irrelevant to consider the case ρ > r. However, PSRG gives a delay from backlog bound; there may be cases where the only information available on the aggregate input is a bound on sustainable rate ρ, with ρ > r. In such cases, there are probably other mechanisms (such as window flow control [47]) to prevent buffer overflow; here, it is useful to be able to bound e as in Theorem 7.5.3. sec 0.05 0.04 0.03 0.02 0.01

50

100

150

200

Mbps

Figure 7.6: Numerical Application of Theorem 7.5.3 and Theorem 2.1.7, showing the additional latency e − e for a composite node, made of a variable delay element (δ = δmax = 0.01s) followed by a PSRG or GR component with rate r = 100Mb/s and latency e. The fresh traffic has arrival curve ρt + σ, with σ = 50KBytes. The figure shows e − e as a function of ρ, for lmin = 0. Top graph: delay element is non-FIFO, second component is PSRG (Theorem 7.5.3); middle graph: delay element is non-FIFO, second component is GR (Theorem 2.1.7); bottom line: delay element is FIFO, both cases (Theorem 7.5.2 and Theorem 7.5.3). Top and middle graph coincide for ρ ≤ r.

sec 0.05

sec 0.05

0.04

0.04

0.03

0.03

0.02

0.02

0.01

0.01

50

100

150

200

Mbps

50

100

150

200

Mbps

Figure 7.7: Same as Figure 7.6, but the fresh traffis has a peak rate limit. The arrival curve for the fresh traffic is min(pt + M T U, ρt + σ), with M T U = 500B, p = 200M b/s (top picture) or p = 2ρ (bottom picture).

7.6. COMPARISON OF GR AND PSRG

209

0.04 0.03 0.02 0.01 0

10000 8000 6000 4000Peak Rate

50 100 Mean rate

2000 150 2000

Figure 7.8: Latency increase as a function of peak rate and mean rate. The parameters are the same as for Figure 7.7. Comment 1 : We now justify why Theorem 7.5.3 is needed, in other words: if we relax the FIFO assumption for the variable delay component, then Theorem 7.5.2 does not hold any more. Intuitively, this is because a tagged packet (say P 3 on Figure 7.5) may be delayed at the second stage by packets (P 4 on the figure) that arrived later, but took over our tagged packet. Also, the service rate may appear to be reduced by packets (P 1 on the figure) that had a long delay in the variable delay component. Formally, we have: P ROPOSITION 7.5.1 (T IGHTNESS ). The bound in Theorem 7.5.3 is tight in the case of an arrival curve of the form α(t) = ρt + σ and if lmax ≥ 2lmin . The proof is in Section 7.7. The proposition shows that the concatenation of non-FIFO PSRG nodes does not follow the rule as for FIFO nodes, which is recalled in the proof of Theorem 7.5.2. Note that if the condition lmax ≥ 2lmin is not satisfied then the bound in Theorem 7.5.3 is tight up to a tolerance of 2lmin /r. Comment 2 : Equation (7.10) for the latency is the minimum of two terms. In the case α(t) = ρt + σ, for ρ ≤ r, the bound is equal to its former term, otherwise to its second term. For a general α however, such a simplification does not occur. Comment 3 : If α is not continuous (thus has jumps at some values), then it can be shown that Theorem 7.5.3 still holds, with Equation (7.10) replaced by e = e + δmax + min{supt≥0 [ α(t+δ) − t], r α0 (t)+α0 (δ) sup0≤t≤δ [ − t]} r with α0 (u) = min[α(u+) − lmin , α(u)].

7.6

C OMPARISON OF GR AND PSRG

First, we know that a PSRG node is GR with the same parameters. This can be used to obtain delay and backlog bounds for arrival curve constrained input traffic. Compare however Theorem 2.1.1 to Theorem 7.3.3: the PSRG characterization has a larger latency e than the GR characterization, so it is better


210

not to use the two characterizations separately: GR to obtain delay and backlog bounds, PSRG to obtain delay-from-backlog bounds. Second, we have shown that for GR there cannot exist a delay-from-backlog bound as in Theorem 7.3.4. Third, there are similar concatenation results as for PSRG in Theorem 2.1.7. The value of latency increase e for the composite node is the same for PSRG and GR when the total incoming rate ρ is less than the scheduler rate r. However, the guarantee expressed by PSRG is stronger than that of GR. Thus the stronger guarantee of PSRG comes at no cost, in that case.

7.7 7.7.1

P ROOFS P ROOF OF L EMMA 7.3.1

In order to simplify the notation, we use, locally to this proof, the following convention: first, ∨ has precedence over ∧; second, we denote A ∨ B with AB. Thus, in this proof only, the expression AB ∧ CD means (A ∨ B) ∧ (C ∨ D) The reason for this convention is to simplify the use of the distributivity of ∨ with respect to ∧ [28], which is here written as A(B ∧ C) = AB ∧ AC Our convention is typical of “min-max” algebra, where min takes the role of addition and max the role of multiplication. Armed with this facilitating notation, the proof becomes simple, but lengthy, calculus. In the rest of the proof we consider some fixed n and drop superscript n . For 0 ≤ j ≤ n − 1, define Fj = fj + mj+1 + ... + mn and let Fn = fn . Also let D0 = d0 + m1 + ... + mn = m1 + ... + mn First note that for all j ≥ 1: fj = (aj + mj ) ∨ [(fj−1 + mj ) ∧ (dj−1 + mj )] then, by adding mj+1 + ... + mn to all terms of the right hand side of this equation, we find Fj = Aj ∨ (Fj−1 ∧ Dj−1 ) or, with our notation: Fj = Aj (Fj−1 ∧ Dj−1 ) and by distributivity: Fj = Aj Fj−1 ∧ Aj Dj−1

(7.11)

Now we show by downwards induction on j = n − 1, ..., 0 that fn =

An An−1 ...Aj+1 Fj ∧ An An−1 ...Aj+1 Dj ∧ ... ∧ An An−1 ...Ak+1 Dk ∧ ... ∧ An An−1 Dn−2 ∧ An Dn−1

(7.12)

7.7. PROOFS

211

where k ranges from j to n − 1. For j = n − 1, the property follows from Equation (7.11) applied for j = n. Assume now that Equation (7.12) holds for some j ∈ {1, ..., n − 1}. By Equation (7.11), we have An An−1 ...Aj+1 Fj = An An−1 ...Aj+1 (Aj Fj−1 ∧ Aj Dj−1 ) thus An An−1 ...Aj+1 Fj = An An−1 ...Aj+1 Aj Fj−1 ∧ An An−1 ...Aj+1 Aj Dj−1 which, combined with Equation (7.12) for j shows the property for j − 1. Now we apply Equation (7.12) for j = 0 and find fn = An An−1 ...A1 F0 ∧ An An−1 ...A1 D0 ∧ ... ∧An An−1 Dn−2 ∧ An Dn−1 First note that F0 = D0 so we can remove the first term in the right hand side of the previous equation. Second, it follows from a1 ≥ 0 that D0 ≤ A1 thus An An−1 ...A1 D0 = An An−1 ...A1 thus finally fn = An An−1 ...A1 ∧ An An−1 ...A2 D1 ∧ ... ∧An An−1 Dn−2 ∧ An Dn−1 which is precisely the required formula.

7.7.2


First, assume that the packet scale rate guarantee holds. Apply Lemma 7.3.1 with mn = for 1 ≤ j ≤ n − 1. fn ≤ max Ann , Ann−1 , ..., Anj+1 , Djn

ln r .

It follows that,

thus fn is bounded by one of the terms in the right hand side of the previous equation. If it is the last term, we have lj+1 + ... + ln fn ≤ Djn = dj + r now dn ≤ fn + e, which shows Equation (7.2). Otherwise, there is some k ∈ {j + 1, ..., n} such that lk + ... + ln r which shows Equation (7.3). For j = 0, Lemma 7.3.1 implies that fn ≤ max Ann , Ann−1 , ..., An1 fn ≤ Ank = ak +

and the rest follows similarly. Second, assume conversely that Equation (7.2) or Equation (7.3) holds. Consider some fixed n, and define Anj , Djn , Fjn as in Lemma 7.3.1, with mn = lrn . For 1 ≤ j ≤ n − 1, we have dn − e ≤ max Ann , Ann−1 , ..., Anj+1 , Djn and for j = 0:

dn − e ≤ max Ann , Ann−1 , ..., An1

thus dn − e is bounded by the minimum of all right-handsides in the two equations above, which, by Lemma 7.3.1, is precisely fn .


212

7.7.3


We first prove that for all i ≥ 0

fi ≥ gi − e1

(7.13)

where fi is defined by Equation (7.1). Indeed, if Equation (7.13) holds, then by Equation (7.5)): di ≤ gi + e2 ≤ fi + e1 + e2 which means that the scheduler offers the packet scale rate guarantee with rate r and latency e = e1 + e2 . Now we prove Equation (7.13) by induction. Equation (7.13) trivially holds for i = 0. Suppose now that it holds for i − 1, namely, fi−1 ≥ gi−1 − e1 By hypothesis, Equation (7.5) holds: di−1 ≥ gi−1 − e1 thus min[fi−1 , di−1 ] ≥ gi−1 − e1

(7.14)

Combining this with Equation (7.1), we obtain fi ≥ gi−1 − e1 +

L(i) R

(7.15)

Again from Equation (7.1) we have fi ≥ ai + lri ≥ ai − e1 +

(7.16)

li ) r

Now by Equation (7.4) gi ≤ max[ai , gi−1 ] +

li r

(7.17)

Combining Equation (7.15)), Equation (7.16)) and (7.17) gives fi ≥ gi − e1

7.7.4


Consider a fixed packet n which is present at time t. Call aj [resp. dj ] the arrival [resp. departure] time of packet j. Thus an ≤ t ≤ dn . Let B be the set of packet numbers that are present in the system at time t, in other words: B = {k ≥ 1|ak ≤ t ≤ dk } The backlog at time t is Q = i∈B li . The absence of FIFO assumption means that B is not necessarily a set of consecutive integers. However, define j as the minimum packet number such that the interval [j, n] is included in B. There is such a j because n ∈ B. If j ≥ 2 then j − 1 is not in B and aj−1 ≤ an ≤ t thus necessarily dj−1 < t (7.18) If j = 1, Equation (7.18) also holds with our convention d0 = 0. Now we apply the alternate characterization of packet scale rate guarantee (Theorem 7.3.2) to n and j − 1. One of the two following equations must hold: lj + ... + ln dn ≤ e + dj−1 + (7.19) r

7.7. PROOFS

213

or there exists a k ≥ j, k ≤ n with dn ≤ e + ak +

lk + ... + ln r

(7.20)

Assume that Equation (7.19) holds. Since [j, n] ⊂ B, we have Qn ≥ lj + ... + ln . By Equation (7.18) and Equation (7.19) it follows that Q dn ≤ e + t + r which shows the result in this case. Otherwise, use Equation (7.20); we have Q ≥ lk + ... + ln and ak ≤ t thus Q dn ≤ e + t + r

7.7.5


Consider some fixed but arbitrary times s ≤ t and let u ∈ [s, t]. We have $ % ˜ − s) ∧ inf [R(v) + β1 (u − v)] R1 (u) ≥ R1 (s) + β(u v∈[s,u]

thus

$ % ˜ − s) + β2 (t − u) ∧ R1 (u) + β2 (t − u) ≥ R1 (s) + β(u inf v∈[s,u] [R(v) + β1 (u − v) + β2 (t − u)]

and inf [R1 (u) + β2 (t − u)] ≥ $ % ˜ − s) + β2 (t − u) inf R1 (s) + β(u

u∈[s,t]

u∈[s,t]

∧

inf

u∈[s,t],v∈[s,u]

[R(v) + β1 (u − v) + β2 (t − u)]

After re-arranging the infima, we find inf [R1 (u) + β2 (t − u)] ≥

$ % ˜ R1 (s) + inf β(u − s) + β2 (t − u) ∧ u∈[s,t]

inf R(v) + inf [β1 (u − v) + β2 (t − u)]

u∈[s,t]

v∈[s,t]

u∈[v,t]

which can be rewritten as inf [R1 (u) + β2 (t − u)] ≥ R1 (s) + (β˜1 ⊗ β2 )(t − s) ∧

u∈[s,t]

inf [R(v) + β(t − v)]

v∈[s,t]

Now by hypothesis we have R∗ (t) ≥ R∗ (s) + β˜2 (t − s) ∧ inf [R(u) + β2 (t − u)] u∈[s,t]


214 Combining the two gives

R∗ (t) ≥ R∗ (s) + β˜2 (t − s) ∧ R1 (s) + (β˜1 ⊗ β2 )(t − s) ∧ inf [R(v) + β(t − v)] v∈[s,t]

Now R1 (s) ≥ R∗ (s) thus R∗ (t) ≥ R∗ (s) + β˜2 (t − s) ∧ R∗ (s) + (β˜1 ⊗ β2 )(t − s) ∧ inf [R(v) + β(t − v)] v∈[s,t]

7.7.6


If the virtual delay at time t is larger than t + τ for some τ ≥ 0, then we must have R∗ (t + τ ) < R(t) By hypothesis

˜ ) ∧ R∗ (t + τ ) ≥ R∗ (t) + β(τ

now for u ∈ [t, t + τ ]

inf

[u∈[t,t+τ ]

(7.21)

[R(u) + β(t + τ − u)]

(7.22)

R(u) + β(t + τ − u) ≥ R(t) + β(0) ≥ R∗ (t + τ )

thus Equation (7.22) implies that

˜ ) R∗ (t + τ ) ≥ R∗ (t) + β(τ

combining with Equation (7.21) gives ˜ ) Q(t) = R(t) − R∗ (t) ≥ β(τ ˜ ) > Q(t)} which is equal to β˜−1 (Q(t)). thus the virtual delay is bounded by sup{τ : β(τ

7.7.7 P ROOF :

P ROOF OF T HEOREM 7.4.4 Let s ≤ t. By hypothesis we have ˜ − s) ∧ inf [R(u) + β(t − u)] R∗ (t) ≥ R∗ (s) + β(t u∈[s,t]

We do the proof when the inf in the above formula is a minimum, and leave it to the alert reader to extend it to the general case. Thus assume that for some u0 ∈ [s, t]: inf [R(u) + β(t − u)] = R(u0 ) + β(t − u0 )

u∈[s,t]

it follows that either or

˜ − s) R∗ (t) − R∗ (s) ≥ β(t R∗ (t) ≥ R(u0 ) + β(t − u0 )

7.7. PROOFS

215

Consider the former case. We have R (t) ≥ R∗ (t) − lmax and R (s) ≤ R∗ (s) thus ˜ − s) − lmax R (t) ≥ R∗ (t) − lmax ≥ R (s) + β(t Now also obviously R (t) ≥ R (s), thus finally ˜ − s) − lmax ] = R (s) + β˜ (t − s) R (t) ≥ R (s) + max[0, β(t Consider now the latter case. A similar reasoning shows that R (t) ≥ R(u0 ) + β(t − u0 ) − lmax but also

R∗ (t) ≥ R(u0 )

now the input is L-packetized. Thus R (t) = P L (R∗ (t)) ≥ P L (R(u0 )) = R(u0 ) from which we conclude that R (t) ≥ R(u0 ) + β (t − u0 ). Combining the two cases provides the required adaptive guarantee.

7.7.8


The first part uses the min-max expansion of packet scale rate guarantee in Lemma 7.3.1. The second part relies on the reduction to the minimum rate server. We use the same notation as in Definition 7.3.1. L(i) = ij=1 lj is the cumulative packet length. I TEM 1: Define the sequence of times fk by Equation (7.1). Consider now some fixed but arbitrary packet index n ≥ 1. By the FIFO assumption, it is sufficient to show that R∗ (t) ≥ L(n)

(7.23)

with t = fn + e. By Lemma 7.3.1, there is some index 1 ≤ j ≤ n such that

L(n) − L(j − 1) + i L(n) − L(k − 1) fn = s + max ak + k=j+1 r r with s = aj ∨ dj−1 and with the convention that d0 = 0. Let us now apply the definition of an adaptive guarantee to the time interval [s, t]: R∗ (t) ≥ A ∧ B with

A := R∗ (s) + r(t − s − e)+ and B := inf B(u) u∈[s,t]

where

' & B(u) := R(u) + r(t − u − e)+

(7.24)

216


Firstly, since s ≥ dj−1 , we have R∗ (s) ≥ L(j − 1). By Equation (7.24), fn ≥ s + t ≥ s + L(n)−L(j−1) + e. It follows that r t−s−e≥

L(n)−L(j−1) r

thus

L(n) − L(j − 1) r

and thus A ≥ L(n). Secondly, we show that B ≥ L(n) as well. Consider some u ∈ [s, t]. If u ≥ an then R(u) ≥ L(n) thus B(u) ≥ L(n). Otherwise, u < an ; since s ≥ aj , it follows that ak−1 ≤ u < ak for some k ∈ {j + 1, ..., n} and R(u) = L(k − 1). By Equation (7.24), fn ≥ ak + thus t−u−e≥

L(n) − L(k − 1) r L(n) − L(k − 1) r

It follows that B(u) ≥ L(n) also in that case. Thus we have shown that B ≥ L(n). Combining the two shows that R∗ (t) ≥ L(n) as required. I TEM 2: We use a reduction to a minimum rate server as follows. Let di := min(di , fi ) for i ≥ 0. By Equation (7.1) we have li (7.25) ai ≤ di ≤ max(ai , di−1 ) + r and di ≤ di ≤ di + e (7.26) The idea of the proof is now to interpret di as the output time for packet i out of a virtual minimum rate server. Construct a virtual node R as follows. The input is the original input R(t). The output is defined as follows. The number of bits of packet i that are output up to time t is ψi (t), defined by ⎧ ⎨ if t > d (i) then ψi (t) = L(i) else if a(i) < t ≤ d (i) then ψi (t) = [L(i) − r(d (i) − t)]+ ⎩ else ψi (t) = 0 so that the total output of R is R1 (t) =

i≥1 ψi (t).

The start time for packet i is thus max[ai , di − lri ] and the finish time is di . Thus R is causal (but not necessarily FIFO, even if the original system would be FIFO). We now show that during any busy period, R has an output rate at least equal to r. Let t be during a busy period. Consider now some time t during a busy period. There must exist some i such that ai ≤ t ≤ di . Let i be the smallest index such that this is true. If ai ≥ di−1 then by Equation (7.25) di − t ≤ lri and thus ψr (t) = r where ψr is the derivative of ψi to the right. Thus the service rate at time t is at least r. Otherwise, ai < D i − 1. Necessarily (because we number packets in order of increasing ai ’s – this is not a FIFO assumption) ai−1 ≤ ai ; since i is the smallest index such that ai ≤ t < di , we must have t ≥ di−1 . But then di − t ≤ lri and the service rate at time t is at least r. Thus, node R offers the strict service curve λr and R → (λr ) → R1 (7.27)

7.7. PROOFS

217

Now define node D. Let δ(i) := di − di , so that 0 ≤ δ(i) ≤ E. The input of D is the output of R. The output is as follows; let a bit of packet i arrive at time t; we have t ≤ di ≤ di . The bit is output at time t = max[min[di−1 , di ], t + δi ]. Thus all bits of packet i are delayed in D by at most δ(i), and if di−1 < di they depart after di . It follows that the last bit of packet i leaves D at time di . Also, since t ≥ t, D is causal. Lastly, if the original system is FIFO, then di−1 < di , all bits of packet i depart after di−1 and thus the concatenation of R and D is FIFO. Note that R is not necessarily FIFO, even if the original system is FIFO. The aggregate output of D is R2 (t) ≥

ψi (t − δ(i)) ≥ R1 (t − e)

i≥1

thus the virtual delay for D is bounded by e and R1 → (δe ) → R2

(7.28)

Now we plug the output of D into an L-packetizer. Since the last bit of packet i leaves D at time di , the final output is R∗ . Now it follows from Equation (7.27), Equation (7.28) and Theorem 7.4.2 that R → (λr ⊗ δe ) → R2

7.7.9


We first introduce some notation (see Figure 7.5). Call an ≥ 0 the arrival times for the fresh input. Packets are numbered in order of arrival, so 0 ≤ a1 ≤ a2 ≤ .... Let ln be the size of packet n. Call bn the arrival time for packet n at the second component; bn is not assumed to be monotonic with n, but for all n: an ≤ bn ≤ an + δ

(7.29)

Also call dn the departure time of packet n from the second component. By convention, a0 = d0 = 0. Then, define e1 = e + δmax + sup[ t≥0

and e2 = e + δmax + sup [ 0≤t≤δ

α(t + δ) − lmin − t] r

α(t) + α(δ) − lmin − t] r

e

= min[e1 , e2 ]. It is sufficient to show that the combined node separately satisfies the packet so that scale rate guarantee with rate r and with latencies e1 and e2 . To see why, define fn by Equation (7.1). If dn − fn ≤ e1 and dn − fn ≤ e2 for all n, then dn − fn ≤ e . Part 1: We show that the combined node satisfies the packet scale rate guarantee with rate r and latency e1 . An arrival curve for the input traffic to the second component is α2 (t) = α(t + δ). Thus, by Theorem 2.1.4, dn ≤ bn + D2 , with α(t + δ) − t] dn ≤ bn + e + sup[ r t≥0 By Equation (7.29): dn − an ≤ e + δmax + sup[ t≥0

α(t + δ) − t] r

Now we apply Lemma 7.5.1 which ends the proof for this part.

218


Part 2: We show that the combined node satisfies the packet scale rate guarantee with rate r and latency e2 . Let δmin = δmax − δ the constant part of the delay. We do the proof for δmin = 0 since we can eliminate the constant delay by observing packets δmin time units after their arrival, and adding δmin to the overall delay. Part 2A: We assume in this part that there cannot be two arrivals at the same instant; in part 2B, we will show how to relax this assumption. For a time interval (s, t] (resp. [s, t]), define A(s, t] as the total number of bits at the fresh input during the interval (s, t] (resp. [s, t]); similarly, define B(s, t] and B[s, t] at the input of the second node. We have the following relations: 1{s0 α(u + ) = α(u) (the last equality is because α is continuous). Second, let l be the packet length for one packet arriving at time t. Then A(t, t + u] + l ≤ A[t, t + u] ≤ α(u) thus A(t, t + u] ≤ α(u) − l ≤ α(u) − lmin . The same reasoning shows the second inequality in the lemma.

Now we apply Theorem 7.3.2. Consider some fixed packets numbers 0 ≤ j < n. We have to show that one of the following holds: A(aj , an ] (7.30) dn ≤ e2 + dj + r or there is some k ∈ {j + 1, ..., n} such that dn ≤ e2 + ak +

A[ak , an ] r

(7.31)

(Case 1:) Assume that bj ≥ bn . Since the second node is FIFO, we have dn ≤ dj and thus Equation (7.30) trivially holds. (Case 2:) Assume that bj < bn . By Theorem 7.3.2 applied to the second node, we have 1 dn ≤ e + dj + B(bj , bn ] r

(7.32)

or there exists some k such that bj ≤ bk ≤ bn and 1 dn ≤ e + bk + B[bk , bn ] r

(7.33)

7.7. PROOFS

219

(Case 2a: ) Assume that Equation (7.32) holds. By Equation (7.29), any packet that arrives at node 2 in the interval (bj , bn ] must have arrived at node 1 in the interval (aj − δ, bn ] ⊂ (aj − δ, an + δ]. Thus B(bj , bn ] ≤ A(aj − δ, an + δ] ≤ A(aj , an ] + A[aj − δ, aj ) + A(an , an + δ] ≤ A(aj , an ] + 2α(δ) − 2lmin the last part being due to Lemma 7.7.1. Thus α(δ) dn ≤ e + δ + α(δ) r − δ + r + dj + 1r A(aj , an ] − 2lmin ≤ e2 + dj + 1r A(aj , an ]

which shows Equation (7.30). (Case 2b: ) Assume that Equation (7.33) holds. Note that we do not know the order of k with respect to j and n. However, in all cases, by Equation (7.29): B[bk , bn ] ≤ A[bk − δ, an + δ]

(7.34)

u = aj − bk + δ

(7.35)

We further distinguish three cases. (Case 2b1: ) k ≤ j: Define By hypothesis, ak ≤ aj and bk − δ ≤ ak so that u ≥ 0. Note also that aj ≤ bj ≤ bk and thus u ≤ δ. By Equation (7.34): B[bk , bn ] ≤ A[bk − δ, aj ) + A[aj , an ] + A(an , an + δ] Now by Lemma 7.7.1 A(an , an + δ] ≤ α(δ) and A[bk − δ, aj ) ≤ α(u) − lmin . Thus B[bk , bn ] ≤ A[aj , an ] + α(u) + α(δ) − 2lmin Combine with Equation (7.33), Equation (7.35) and obtain dn ≤ aj +

A[aj , an ] + e2 r

which shows that Equation (7.31) holds. (Case 2b2: ) j < k ≤ n: Define u = δ − bk + ak . By Equation (7.34) B[bk , bn ] ≤ A[ak , an ] + α(u) + α(δ) − 2lmin which shows that

1 dn ≤ e2 + ak + A[ak , an ] r

(Case 2b3: ) k > n: Define u = δ − bk + an . By bk ≤ bn and bn ≤ an + δ we have u ≥ 0. By bk ≥ ak and ak ≥ an we have u ≤ δ. Now by Equation (7.33): 1 1 dn ≤ e + bk + B[bk , bn ] = e + δ − u + an + B[bk , bn ] r r


220 By Equation (7.34)

B[bk , bn ] ≤ A[an − u, an + δ] = A[an − u, an ) + ln + A(an , an + δ] ≤ α(u) + ln + α(δ) − 2lmin which shows that dn ≤ e2 + an +

ln r

Part 2B: Now it remains to handle the case where packet arrivals at either component may be simultaneous. We assume that packets are ordered at component 2 in order of arrival, with some unspecified mechanism for breaking ties. Packets also have a label which is their order of arrival at the first component; we call (k) the label of the kth packet in this order (see Figure 7.5 for an illustration). Call S the original system. Fix some arbitrary integer N . Consider the truncated system S N that is derived from the original system by ignoring all packets that arrive at the first component after time aN + δ. Call N N N aN n , bn , dn , fn the values of arrival, departure, and virtual finish times in the truncated system (virtual finish times are defined by Equation (7.1)). Packets with numbers ≤ N are not affected by our truncation, N N N thus aN n = an , bn = bn , dn = dn , fn = fn for n ≤ N . Now the number of arrival events at either component 1 or 2 in the truncated system is finite; thus we can find a positive number η which separates arrival events. Formally: for any m, n ≤ N : am = an or |am − an | > η and bm = bn or |bm − bn | > η Let < η2 . We define a new system, called S N, , which is derived from S N as follows. • We can find some sequence of numbers xn ∈ (0, ), n ≤ N such that: (1) they are all distinct; (2) if the packet labeled m is ordered before the packet labeled n in the order of arrival at the second component, then xm < xn . Building such a sequence is easy, and any sequence satisfying (1) and (2) will do. For example, take xn = Nk+1 where k is the order of arrival of packet n (in other words, (k) = n). • Define the new arrival and departure times by a n = an + xn , b n = bn + xn , d n = dn + xn It follows from our construction that all a n are distinct for n ≤ N , and the same holds for b n . Also, the arrival order of packets at the second component is the same as in the original system. Thus we have built a new system S N, where all arrivals times are distinct, the order of packets at the second component is the same as in S N , arrival and departure times are no earlier than in S N , and differ by at most .

the virtual finish times at the second component. By definition: For k ≤ N , call F(k)

⎧ ⎪ F = 0 ⎪ % $ ⎨ (0)

= max b , min d

F(k) , F (k) (k−1) (k−1) ⎪ ⎪ ⎩ + l(k) for k ≥ 1 r

and a similar definition holds for F(k) by dropping . It follows by induction that

≥ F(k) F(k)


221

thus

d (k) ≤ dk + ≤ e + F(k) ≤ e + F(k) +

Similarly, b k ≤ a k + δ. This shows that S N, satisfies the assumptions of the theorem, with e replaced by e+ Thus the conclusion of Part 2A holds for S N, . Define now fn by Equation (7.1) applied to a n and d n . We have: d n ≤ fn + e2 +

(7.36)

It also follows by induction that fn ≤ fn + Now dn ≤ d n thus dn − fn ≤ d n − fn + Combining with Equation (7.36) gives: dn − fn ≤ e2 + 2 Now can be arbitrarily small, thus we have shown that for all n ≤ N : dn − fn ≤ e2 Since N is arbitrary, the above is true for all n.

7.7.10

P ROOF OF P ROPOSITION 7.5.2

Proof. Case ρ ≤ r: Assume that the source is greedy from time 0, with packet n = 1, of size l1 = lmin , a1 = 0, b1 = δmax . Assume all subsequent packets have a delay in the first component equal to δmax − δ. We can build an example where packet 1 is overtaken by packets n = 2, ..., n1 that arrive in the interval (0, δ], with l2 +...+ln1 = ρδ+σ−l1 . Assume that packet 1 undergoes . Now the maximum delay allowed by PSRG at the second component. It follows after some algebra that d1 = e + δmax + ρδ+σ r f1 = lmin thus d − f = e and the characterization is tight. 1 1 r Case ρ > r: We build a worst case scenario as follows. We let e = 0, without loss of generality (add a delay element to this example and obtain the general case). The principle is to first build a maximum size burst which is overtaken by a tagged packet j. Later, a tagged packet n is overtaken by a second maximum size burst. Between the two, packets arrive at a data rate r; the second burst is possible because r < ρ and an − aj is long enough. Details are in Figure 7.9 and Table 7.1. We find finally dn − fn = 2(ρδ + σ − lmin )/r which shows that the bound is achieved.

7.8


The concept of adaptive service curve was introduced in Okino’s dissertation in [62] and was published by Agrawal, Cruz, Okino and Rajan in [1], which contains most results in Section 7.4.2, as well as an application to a window flow control problem that extends Section 4.3.2 on Page 147. They call β˜ an “adaptive service curve” and β a “partial service curve”. The packet scale rate guarantee was first defined independently of adaptive service guarantees in [4]. It serves as a basis for the definition of the Expedited Forwarding capability of the Internet.


222

a P

n o n F I F O , d e la y d k

F I F O , P S R G

b

1

k

2

d k

1

0

P

P d

P 0

j j + 1

j

P 1

P

d

P j + 1

j

P

P a

a

+ d n

a

P n

a n

k

P

n + 1

n + 1

n

a

1

P n

+ d

b k

P

j + 1

n

d k

Figure 7.9: Worst-case example for Theorem 7.5.3. All packets have 0 delay through the first component except packets 1...j − 1 and n.

k ak lk bk fk dk + 1 0 σ − lmin δ not relevant dj + l1 /r 2 l2 /ρ l2 δ+ not relevant dj + (l1 + l2 )/r ... ... ... ... ... ... j−1 δ lj−1 δ+ not relevant dj + A j δ lmin δ ≥ δ + lmin /r δ + lmin /r j+1 δ + lmin /r lmin aj+1 δ + 2lmin /r fj+1 + A ... ... ... ... ... ... n−1 δ + (n − j − 1)lmin /r lmin an−1 δ + (n − j)lmin /r fn−1 + A n δ + (n − j)lmin /r lmin an + δ δ + (n − j + 1)lmin /r fn + 2A + n+1 an σ − lmin an+1 not relevant fn−1 + A + (σ − lmin )/r n+2 an + a2 l2 an+2 not relevant fn−1 + A + (σ − lmin + l2 )/r ... ... ... ... ... ... n+j−1 (an + δ)− lj−1 (an + δ)− not relevant fn−1 + 2A Notes: A = (ρδ + σ − lmin )/r ρδ (j, l2 , ..., lj−1 ) is a solution to l2 + ... + lj−1 = ρδ, sc l2 , ..., lj−1 ∈ [lmin , lmax ]. For example, let j = 2 + lmin , l2 = ρδ − (j − 3)lmin , l3 = ... = lj−1 = lmin . We have l2 ≤ lmax because lmax ≥ 2lmin Table 7.1: Details for Figure 7.9. Assume for this table that σ − lmin ≤ lmax , otherwise replace packets 1 and n + 1 by a number of smaller packets arriving in batch.

7.9. EXERCISES

7.9

223

E XERCISES

˜ β) → R∗ . E XERCISE 7.1. Assume that R → (β, 1. Show that the node offers to the flow a strict service curve equal to β˜ ⊗ β, where β is the sub-additive closure of β. 2. If β˜ = β is a rate-latency function, what is the value obtained for the strict service curve ? E XERCISE 7.2. Consider a system with input R and output R∗ . We call “input flow restarted at time t” the flow Rt defined for u ≥ 0 by Rt (u) = R(t + u) − R∗ (t) = R(t, u] + Q(t) where Q(t) := R(t) − R∗ (t) is the backlog at time t. Similarly, let the“output flow restarted at time t” be the flow Rt∗ defined for u ≥ 0 by Rt∗ (u) = R∗ (t + u) − R∗ (t) Assume that the node guarantees a service curve β to all couples of input, output flows (Rt , Rt∗ ). Show that R → (β) → R∗ .

224


C HAPTER 8

T IME VARYING S HAPERS

8.1

I NTRODUCTION

Throughout the book we usually assume that systems are idle at time 0. This is not a limitation for systems that have a renewal property, namely, which visit the idle state infinitely often – for such systems we choose the time origin as one such instant. There are cases however where we are interested in the effect at time t of non zero initial conditions. This occurs for example for re-negotiable services, where the traffic contract is changed at periodic renegotiation moments. An example for this service is the Integrated Service of the IETF with the Resource reSerVation Protocol (RSVP), where the negotiated contract may be modified periodically [33]. A similar service is the ATM Available Bit Rate service (ABR). With a renegotiable service, the shaper employed by the source is time-varying. With ATM, this corresponds to the concept of Dynamic Generic Cell Rate Algorithm (DGCRA).. At renegotiation moments, the system cannot generally be assumed to be idle. This motivates the need for explicit formulae that describe the transient effect of non-zero initial condition. In Section 8.2 we define time varying shapers. In general, there is not much we can say apart from a direct application of the fundamental min-plus theorems in Section 4.3. In contrast, for shapers made of a conjunction of leaky buckets, we can find some explicit formulas. In Section 8.3.1 we derive the equations describing a shaper with non-zero initial buffer. In Section 8.3.2 we add the constraint that the shaper has some history. Lastly, in Section 8.4, we apply this to analyze the case where the parameters of a shaper are periodically modified. This chapter also provides an example of the use of time shifting.

8.2

T IME VARYING S HAPERS

We define a time varying shaper as follows. D EFINITION 8.2.1. Consider a flow R(t). Given a function of two time variables H(t, s), a time varying shaper forces the output R∗ (t) to satisfy the condition R∗ (t) ≤ H(t, s) + R∗ (s) for all s ≤ t, possibly at the expense of buffering some data. An optimal time varying shaper, or greedy time varying shaper, is one that maximizes its output among all possible shapers. 225

CHAPTER 8. TIME VARYING SHAPERS

226

The existence of a greedy time varying shaper follows from the following proposition. P ROPOSITION 8.2.1. For an input flow R(t) and a function of two time variables H(t, s), among all flows R∗ ≤ R satisfying R∗ (t) ≤ H(t, s) + R∗ (s) there is one flow that upper bounds all. It is given by R∗ (t) = inf H(t, s) + R(s) s≥0

(8.1)

where H is the min-plus closure of H, defined in Equation (4.10) on Page 142. P ROOF :

The condition defining a shaper can be expressed as ∗ R ≤ LH (R∗ ) R∗ ≤ R

where LH is the min-plus linear operator whose impulse response is H (Theorem 4.1.1). The existence of a maximum solution follows from Theorem 4.3.1 and from the fact that, being min-plus linear, LH is upper-semi-continuous. The rest of the proposition follows from Theorem 4.2.1 and Theorem 4.3.1. The output of the greedy shaper is given by Equation (8.1). A time invariant shaper is a special case; it corresponds to H(s, t) = σ(t − s), where σ is the shaping curve. In that case we find the well-known result in Theorem 1.5.1. In general, Proposition 8.2.1 does not help much. In the rest of this chapter, we specialize to the class of concave piecewise linear time varying shapers. P ROPOSITION 8.2.2. Consider a set of J leaky buckets with time varying rates rj (t) and bucket sizes bj (t). At time 0, all buckets are empty. A flow R(t) satisfies the conjunction of the J leaky bucket constraints if and only if for all 0 ≤ s ≤ t: R(t) ≤ H(t, s) + R(s) with

H(t, s) = min {bj (t) + 1≤j≤J

t

rj (u)du}

(8.2)

s

Consider the level of the jth bucket. It is the backlog of the variable capacity node (SecP ROOF : tion 1.3.2) with cumulative function t rj (u)du Mj (t) = 0

We know from Chapter 4 that the output of the variable capacity node is given by Rj (t) = inf {Mj (t) − Mj (s) + R(s)} 0≤s≤t

The jth leaky bucket constraint is

R(t) − Rj (t) ≤ bj (t)

Combining the two expresses the jth constraint as R(t) − R(s) ≤ Mj (t) − Mj (s) + bj (t) for all 0 ≤ s ≤ t. The conjunction of all these constraints gives Equation (8.2). In the rest of this chapter, we give a practical and explicit computation of H for H given in Equation (8.2), when the functions rj (t) and bj (t) are piecewise constant.

8.3. TIME INVARIANT SHAPER WITH INITIAL CONDITIONS

8.3

227

T IME I NVARIANT S HAPER WITH N ON - ZERO I NITIAL C ONDITIONS

We consider in this section some time invariant shapers. We start with a general shaper with shaping curve σ, whose buffer is not assumed to be initially empty. Then we will apply this to analyze leaky bucket shapers with non-empty initial buckets.

8.3.1

S HAPER WITH N ON - EMPTY I NITIAL B UFFER

P ROPOSITION 8.3.1 (S HAPER WITH NON - ZERO INITIAL BUFFER ). Consider a shaper system with shaping curve σ. Assume that σ is a good function. Assume that the initial buffer content is w0 . Then the output R∗ for a given input R is R∗ (t) = σ(t) ∧ inf {R(s) + w0 + σ(t − s)} 0≤s≤t

for all t ≥ 0

(8.3)

P ROOF : First we derive the constraints on the output of the shaper. σ is the shaping function thus, for all t≥s≥0 R∗ (t) ≤ R∗ (s) + σ(t − s) and given that the bucket at time zero is not empty, for any t ≥ 0, we have that R∗ (t) ≤ R(t) + w0 At time s = 0, no data has left the system; this is expressed with R∗ (t) ≤ δ0 (t) The output is thus constrained by R∗ ≤ (σ ⊗ R∗ ) ∧ (R + w0 ) ∧ δ0 where ⊗ is the min-plus convolution operation, defined by (f ⊗ g)(t) = inf s f (s) + g(t − s). Since the shaper is an optimal shaper, the output is the maximum function satisfying this inequality. We know from Lemma 1.5.1 that R∗ = σ ⊗ [(R + w0 ) ∧ δ0 ] = [σ ⊗ (R + w0 )] ∧ [σ ⊗ δ0 ] = [σ ⊗ (R + w0 )] ∧ σ which after some expansion gives the formula in the proposition.

.

Another way to look at the proposition consists in saying that the initial buffer content is represented by an instantaneous burst at time 0. The following is an immediate consequence. C OROLLARY 8.3.1 (BACKLOG FOR A SHAPER WITH NON - ZERO INITIAL BUFFER ). The backlog of the shaper buffer with the initial buffer content w0 is given by w(t) = (R(t) − σ(t) + w0 ) ∨ sup {R(t) − R(s) − σ(t − s)} 00

with σ 0 (u) = min (rj · u + bj − qj0 ) 1≤j≤J

for all t ≥ 0

(8.6)

8.4. TIME VARYING LEAKY-BUCKET SHAPER P ROOF :

229

Apply Proposition 8.3.3 to the input R = (R + w0 ) ∧ δ0 and observe that σ 0 ≤ σ.

An interpretation of Equation (8.6) is that the output of the shaper with non-zero initial conditions is either the output of the ordinary leaky-bucket shaper, taking into account the initial level of the buffer, or, if smaller, the output imposed by the initial conditions, independent of the input.

8.4

T IME VARYING L EAKY-B UCKET S HAPER

We consider now time varying leaky-bucket shapers that are piecewise constant. The shaper is defined by a fixed number J of leaky buckets, whose parameters change at times ti . For t ∈ [ti , ti+1 ) := Ii , we have thus rj (t) = rji and bj (t) = bij At times ti , where the leaky bucket parameters are changed, we keep the leaky bucket level qj (ti ) unchanged. We say that σi (u) := min1≤jJ {rji u + bij } is the value of the time varying shaping curve during interval Ii . With the notation in Section 8.2, we have H(t, ti ) = σi (t − ti ) if t ∈ Ii We can now use the results in the previous section. P ROPOSITION 8.4.1 (B UCKET L EVEL ). Consider a piecewise constant time varying leaky-bucket shaper with output R∗ . The bucket level qj (t) of the j-th bucket is, for t ∈ Ii : $ % qj (t) = R∗ (t) − R∗ (ti ) − rji · (t − ti ) + qj (ti ) ∨ (8.7) supti 0} dN (t) = 0 (9.29) 0∞ 1{W (t) 0 n−1 sup sup |z(si ) − z(si+1 )| < ∞. n∈N0 0=sn